tech

2025-09-28 · open the original ↗

• Whether a job "changes" or "disappears" matters little to the worker. Frey notes that laundresses were automated by the washing machine, not by a robot copying their motions, so judging automatability from a job's task list says little. A self-driving school bus may still need a minder, but that person needs no driver's licence and has a different skill set, so the driver is replaced either way.
• Automation alone yields little growth; new industries do. Frey argues that if everything since 1800 had been automation, we would have cheap textiles and productive farming but no vaccines, antibiotics, aircraft or computers. Most work done in the US today did not exist in 1940. He warns that today's new firms create fewer jobs than they once did, and that AI used mainly as a productivity tool will bring a short-term boost but little lasting growth or job creation.
• One-person unicorns would not deliver shared prosperity. Answering the host's citation of Sam Altman's prediction of a one-person billion-dollar company, Frey says such firms could be taxed to fund services. Unless their services spread through the whole economy, though, a few unicorns produce neither broad prosperity nor much productivity growth.
• AI needs new sectors, as railways and cars once were. Frey says Britain's growth took off only with the railways, after about seven decades centred on mechanising textiles. The second industrial revolution added cars, electrical goods, component and machine-tool industries, road commerce and tourism. The computer revolution produced a smaller surge that faded quickly, and he expects AI to follow that path unless it creates comparably large sectors.

2025-09-27 · open the original ↗

• DevOps got its vocabulary once isolated practitioners finally met. Jacob, CEO and co-founder of System Initiative and co-founder of Chef, says people independently built CI, deployment and configuration-management systems (his first in 1998) and only named them continuous delivery or DevOps after meeting at O'Reilly's Velocity conference. He credits Mark Burgess, author of CFEngine and promise theory, with writing down why the approach works.
• Infrastructure as code works, but most practitioners hate it. Jacob likens it to Dark Souls: mastering it feels great, which convinces the masters it was good, but for most people it means slow feedback, fragile pieces and compliance failures found in an audit six months after the change. His example of what runs underneath is a bank whose Brazilian reconciliation-law compliance depended on undeployable Delphi software on a Windows machine under someone's desk.
• LLMs are creative, not intelligent, Jacob argues. He defines creativity as making something that did not exist before, such as a sonnet no program could write. Hanselman objects that next-token prediction is Mad Libs and that persistence is not creativity; Jacob concedes it is Mad Libs at the brute-force level and cites an AI topographic map of Washington DC that looks right but is entirely misplaced, because nothing deterministic grounded it.
• AI needs a deterministic layer that gives feedback immediately. Jacob says "rubbing an LLM all over" existing systems fails; the less the model is asked to do, and the sooner it learns whether a proposal is valid (at the moment of change, not the end of a PR cycle), the better it performs. Abstraction layers over AWS or Azure confuse it, while a model that maps one-to-one to the providers' own documentation gets it right "almost all the time."
• System Initiative turns infrastructure from source code into data. Answering Hanselman's point that prose cannot be regenerated exactly, Jacob says intent updates a data model that can change real resources, generate templates or duplicate an app into another region with one change. His reported examples from his own company: an LLM found an off-by-one-character configuration parameter in minutes, discovered infrastructure from a drawn diagram, and migrated EC2 workloads to containers; the free tier is 100 resources.

2025-09-27 · open the original ↗

• LLMs recombine their training data well but cannot make inconceivable leaps. Frey cites a thought experiment by two Oxford colleagues: an LLM trained on data up to 1900 would see only failed flight attempts and no bird over 30 pounds that gets off the ground, so it would not have produced the Wright brothers' blueprint. A model trained on Impressionist paintings yields near-Impressionism, not conceptual art. For frontier discovery, he concludes, humans are safe "for now", while stressing he is not a computer scientist.
• AI still awaits its separate-condenser moment. Frey compares today's AI to early steam engines, too energy-inefficient to do more than drain coal mines (and badly) until the separate condenser and later tweaks made them useful. Brute force works only on a static distribution, he argues, and the world keeps changing. He cites a 2023 paper from Stuart Russell's lab showing amateurs on standard laptops beating superhuman Go programs by exposing them to new concepts.

2025-09-26 · open the original ↗

• In law, AI mainly augments lawyers rather than replacing them. The host names contract review, legal research and document automation as the commonest uses, plus some litigation-outcome prediction. He says they save attorneys hundreds of hours a year and free them for strategy, advocacy and client advice, and that hybrid "legal technologist" roles are emerging.
• Paralegal work will shrink but paralegals need not be left behind. The host concedes that document review, discovery and due diligence are exactly what AI now automates. He says firms are retraining paralegals to supervise AI, validate outputs and manage data workflows. Compliance, data privacy and AI regulation are opening new roles, and he argues entry-level lawyers who adapt may be in greater demand than before.
• Funded startups are driving the change. The host describes Harvey, backed by OpenAI, as a natural-language co-pilot that major firms have rolled out to thousands of lawyers. He reports that Thomson Reuters spent $650 million acquiring the company behind CoCounsel, a research, deposition-prep and contract-analysis tool.

2025-09-25 · open the original ↗

• Copilot Chat's three modes each add power to the one before. Giard says Ask mode only answers questions and leaves code alone, Edit mode changes code, and Agent mode can also run commands for you. He works the intermediate JavaScript lesson of the Mastering GitHub Copilot course from GitHub and Microsoft Learn in its prebuilt Codespace, browser-based Visual Studio Code with the recipe app already installed. Typing a slash lists the built-in commands: /doc wrote documentation for every get and post method in the open routes.js, and /help explains the rest.
• A plain code comment is enough to generate a new route. Typing a // comment asking for a route that gets a random recipe made Copilot propose code in lighter italic text, and Tab accepted it. The result was a recipe/random route that connects to the database and runs select * from recipes order by random limit 1. Copilot then suggested a delete-recipe route he had not asked for, and Escape turned it down.
• Inline chat edits only the code you highlight. With a route highlighted, Ctrl-I (Command-I on a Mac) opens a small chat in the editor. "Sort by title" added order by title to the query and showed the change in green. From there you can accept it, regenerate it for another attempt, or close it to throw it away.
• Agent mode carried out a multi-file feature, tests included. His prompt asked it to delete recipes through an HTML form and a new POST route on the recipe ID, put buttons on the view and edit pages, and write tests. Giard calls it a good prompt because it gives everything he knows without saying how to do it. In under a minute it changed routes.js, added a test, and added a delete button to a page. He kept each change after reviewing it one at a time, then allowed it to run the tests, and three tests passed.
• Instruction files make Copilot follow your project's conventions. Copilot already tries to match the style of the existing files, but copilot-instructions.md in the .github folder states it outright for the whole project: here Node.js with Express, SQL.js, Handlebars, Jest, folder locations for queries and views, and coding patterns. Files named *.instructions.md in an instructions folder cover only part of the project; the styling file's header limits it to the CSS files in public. To test it he asked for a new FAQ view, and Copilot created a Handlebars .hbs file in the views folder, as the instructions specify.

2025-09-25 · open the original ↗

• Edit mode changes code across several files at once. Ask mode, used in steps 1 and 2, only answered questions. In edit mode Copilot shows its changes inline, and nothing is applied until you keep it. Giard says working across files is the advantage because a change often touches more than one.
• Adding files to the context tells Copilot where to look. You can pick files and folders with the Add Context button in Copilot Chat. Giard finds it simpler to drag them into the chat window, here the JS, HTML and CSS files.
• One prompt added participant lists to the activity cards. The activity data already held an array of participants, such as two for the chess club, but the cards did not show it. Giard asked Copilot to add a participants section as a bulleted list and "make it pretty". It built the list's HTML in app.js, appended it to the card, and added styles in styles.css. Each of the four changes could be kept or undone. After he kept them, the cards listed participants inside a border with an offset colour.

2025-09-25 · open the original ↗

• Ask mode with @workspace traced a bug without touching code. The bug was that a student could sign up twice for the same activity, or for more than one; Giard wants one activity per student. Copilot searched all 24 files in the repository and placed the bug in the signup function in app.py. It recommended checking whether the email is already in the activity's participants.
• A plain-language comment got Copilot to write the fix. Giard typed the comment "validate student is not already signed up" and accepted the greyed-out suggestions with Tab. The new code checks every activity for the email and raises an exception if it is there. On retest, a second signup was refused with "Student already signed up for an activity".
• Inline chat (Ctrl+I) works on the code under the cursor. On Mac it is Cmd+I. With the cursor on the activities collection, Giard asked for two more sports, two more artistic and two more intellectual activities, and Copilot extended the list. You can accept, close to reject, or regenerate for a different version. The new activities appeared both in the cards and in the drop-down.
• Copilot drafts the commit message from the staged changes. Giard says a relevant message lets a later reader see when a change happened. Writing one used to mean reviewing the diff himself. The sparkle button above the commit button in Source Control writes a message from the changes, which he says you should still read and check. Committing first asked to stage the changed file.

2025-09-25 · open the original ↗

• GitHub and Microsoft Learn's course comes as hands-on exercise repositories. The course is in GitHub's Microsoft organization, as "mastering GitHub Copilot for paired programming". Its lessons are grouped as beginner, intermediate and advanced. Giard recommends forking the repository. Copy Exercise then puts a lesson's source and instructions in your own repository. There, a GitHub agent called "Professor Cat" watches your progress and gives tips.
• Codespaces runs VS Code and its extensions in the browser. This avoids installing an IDE, Python and extensions locally, and having to remove them afterwards; any IDE with the right extensions also works. The lesson needs the GitHub Copilot and Python extensions. Giard recommends VS Code itself: free, cross-platform, open source and lightweight. He flags an error in the instructions: his environment had no Copilot icon, but View > Chat in the hamburger menu reaches Copilot Chat.
• Ask mode with @workspace explains an unfamiliar codebase. Of the three modes (ask, agent, edit), ask answers without changing code. @workspace understands the whole repository. Asked to explain the structure, it read 24 files and described the folders, configuration and key files. Giard says that as a consultant who inherited other people's code, this would have got him up to speed quickly.
• Copilot supplies setup commands you can send straight to the terminal. A follow-up, "how do I run it?", kept the context. Copilot gave commands to create a Python virtual environment, upgrade pip, install from requirements.txt, and start the app with uvicorn on port 8000. Each answer can be copied, inserted at the cursor, or inserted into the terminal. The app lets a student sign up by email for one of three activities.
• General Git questions get the commands, too. Asked how to create and publish a new branch, Copilot gave a checkout command to create the branch and a push. Giard ran them from the chat and the new branch appeared. He pitches this for tools you use too rarely to remember the syntax, instead of a search engine.

2025-09-25 · open the original ↗

• Innovation depends on freedom to choose collaborators, not on regime type. Frey takes "weak ties" from a 1970s sociology paper and the "collective brain" from Joseph Henrich. Katalin Karikó left socialist Hungary and struggled for decades for US funding, was denied tenure, but found collaborators and laid the groundwork for mRNA vaccines built within months of the virus being identified. Page and Brin met at Stanford, and parents in science and the density of inventors where one grows up are strong predictors of becoming an inventor. Elite institutions in Russia and China foster such ties too, he says.
• Incumbents past the easy gains stop innovating and start lobbying. Once process improvements hit diminishing returns, Frey argues, protecting oneself from competition pays better, and incumbents avoid breakthroughs that would cannibalise their business. His examples are Kodak shelving its early digital camera and Google, whose transformer authors left and which did not launch generative AI before OpenAI, perhaps to protect ad revenue. He cites research showing creative talent clustering in a few incumbents, where inventors are less productive than at their earlier startups.
• Antitrust, not coordination, gave America the computer revolution. Frey says Japan led in autos, semiconductors and consumer electronics in the 1980s, then stagnated while US productivity surged around 1995. He credits the IBM case forcing hardware and software apart, which let Microsoft enter, and the AT&T break-up, which kept ARPANET from being bottlenecked in one boardroom. Japan's keiretsu, by contrast, raised barriers to entry. He also notes US business dynamism is in decline and draws the lesson for US competition with China.

2025-09-24 · open the original ↗

• Giard records GCast as a one-person Microsoft Teams meeting. He uses Teams because it is already installed; he used to use TechSmith's Camtasia, but doesn't have the latest version on this machine. He starts a Meet Now meeting with the camera off and shares a single monitor, keeping his notes on another. He then starts "record and transcribe", and when done stops recording and ends the meeting.
• The recording downloads from the browser, not from Teams. The meeting recap in Teams holds the recording and transcript. Giard found no download option there, so he uses "watch in browser", where a download button saves an MP4.
• Each episode starts as a copy of the last Premiere project. He edits in Adobe Premiere Pro because he has it and wants to get better at it. Copying the last project keeps the intro, outro and music. He saves it under the new episode number and title, deletes the old screencast, and updates the episode number, title and recording date. The raw MP4 goes in a per-year raw-videos folder under a GCast videos folder on his C drive. From there he imports it with the Media Browser and drags it onto the timeline. He trims the start to where the audio begins, using Ctrl+K and Delete.
• Zooming the clip hides the Teams self-view thumbnail. He knows no way to remove the thumbnail itself. In Effect Controls he sets X position 1155, Y position 562 and scale 120, numbers he found by experiment, which fills the frame with a little cropping.
• He exports to a final-output folder, then uploads to YouTube. The file is named "GCast" plus the episode number and a description. For a project without a template, assets are imported through the Media Browser and dragged onto the timeline. These include the intro animation he made in PowerPoint a couple of years ago and Creative Commons music credited at the end. A text tool adds titles. Giard says the template takes out the tedious, repetitive part. The fun part is building the demos.

2025-09-24 · open the original ↗

• Agents act on their own; assistants only guide you. Gomez, head of AI at a medium-sized consulting company, contrasts agents with traditional automation such as console apps or robotic process automation (RPA), which follow structured flows for every scenario you can anticipate; an agent takes action from its prompt and the tools you give it, with as much independence as you choose.
• Document-heavy, low-judgment work is the ideal agent use case. His example is processing invoices or other PDFs, extracting information and then acting on it, such as creating a service ticket or entering data in another database: work humans do now that needs little human intelligence.
• Azure AI Foundry's agent service is his recommended no-code start. He says you can build agents in the portal without code, add MCP servers, combine them with Azure Functions or Logic Apps, and inspect each agent's threads and actions to adjust prompts and tools; SDKs let you call the same agents from code later. He names LangGraph as an orchestration option for teams already on LangChain.
• Use an agent only where simpler automation won't do. He compares the AI rush to microservices, from which some organizations are now moving back to monoliths; pick a business scenario with good ROI first, and prefer a traditional approach for "if this then that" logic, where LLM costs are hard to estimate compared with counted API calls, or where results must be identical every time.
• Plan resources for fast change and for responsible use. He warns that Foundry SDKs are likely to change quickly with breaking changes that can stop agents working, that the field is too new for Stack Overflow to help (he points to Microsoft's docs and a Foundry Discord), and that solutions need content-safety safeguards and care not to expose sensitive data.

2025-09-24 · open the original ↗

• Progress has two phases, exploration and scaling, which reward different systems. Frey likens exploration to a startup finding its model and scaling to low-cost delivery of what is already known. A Soviet aircraft engineer could ask the Red Army and at most two or three others, and needed his idea to match national priorities. In the US, Bessemer Venture Partners passing on Google in 1999 did not kill it, because Sequoia took the bet. More funders explore more trajectories, so he expects breakthroughs from decentralised systems.
• A system built for scaling broke when computing demanded exploration. Frey says the Soviet Union grew fast for decades by mobilising resources in heavy industry such as steel. Geography, institutions and culture do not explain its 1991 collapse, he argues: geography was unchanged, institutions were arguably better under Gorbachev's glasnost and perestroika, and Western broadcasts should have made the culture more progressive. The explanation he offers is the interaction of institutions and culture with changing technology.

2025-09-23 · open the original ↗

• "Hacker" describes a way of thinking, not a criminal. Street rejects "white hat" and "black hat" as stigmatizing, arguing that nobody asks for an ethical dentist; he counts artists, inventors and scientists as hackers because they try what the world says is not done.
• Social engineering works by keeping targets on autopilot. He makes a situation seem only slightly odd, such as claiming to run a "USB audit" from headquarters. In a bank he robbed abroad (on video) he compromised the first machine within 15 seconds; a teller who rightly refused passed him to a manager without briefing him, each assumed the other had vetted him, and the manager escorted him to every machine in the building.
• Discomfort and sympathy beat promises of prizes. In a password-reset video check he says he posed as a tearful man packing up his dead stepfather's house, and the help desk reset the password, because people want off a call about death or illness. He calls a help-desk worker who held to policy for 40 minutes one of his best wins.
• The real work is turning the breach into a lesson. He says he mimics an adversary so staff remember what one looks like, and tells them "I wasn't testing you, I was teaching you"; he puts positive reactions at about 99% overall and almost 100% in person. He now uses AI for convincing fake websites that pass web filters, phishing pretexts, local colloquialisms for blending in, and voice cloning and deepfakes built from a single CEO interview.
• Treat any odd request as a "child in the road." He argues that, as with driving on autopilot, the brain flags anomalies, and you should then stop and verify out of band: put the caller on hold, contact the sender directly, and blame management's policy to stay polite. Against deepfaked calls he recommends analog code words, like the laminated card of 12 monthly words his former bank mailed clients each December for wires over $100,000.

2025-09-23 · open the original ↗

• Breakthroughs need decentralised, competitive systems; progress is not inevitable. In How Progress Ends, Frey notes few leaders (Song China, the Dutch Republic, Victorian Britain) stayed ahead long. Centralised economies scale but explore poorly: a Soviet engineer had the Red Army and two or three other funders, while Sequoia backed Google after Bessemer passed in 1999. He credits the IBM unbundling case and the AT&T break-up for the US software and internet boom from about 1995, while Japan's keiretsu raised barriers to entry. Incumbents past easy gains turn to lobbying, as with Kodak's digital camera and Google's transformer work.
• LLMs are engines of statistical consensus, not invention. Frey argues a model trained on data to 1900 would have seen only failed flights and no bird over 30 pounds that flies, so it would not have foreseen the Wright brothers. He says AI awaits its "separate condenser" moment, which made early steam engines useful. He cites a 2023 paper from Stuart Russell's lab in which amateurs on laptops beat superhuman Go programs with novel concepts. He also calls the latest models better tutors than he is, says their gains will soon be imperceptible to users, and puts the human edge in applying the right model to the right problem.
• AI used only for automation will bring little growth or jobs. Frey notes most US work today did not exist in 1940, new firms now create fewer jobs, and growth came from new sectors: railways, then cars, electrical goods and tourism. The computer revolution's surge faded fast, and he expects the same of AI unless it builds comparable industries. A one-person billion-dollar firm, the Sam Altman prediction the host cited, would be taxable but not shared prosperity.
• AI erodes white-collar jobs mainly by lowering barriers and enabling offshoring. Frey compares it to GPS making taxi knowledge worthless. Since a New York accountant earns six figures against a low four-digit salary in Manila, he expects narrower productivity gaps to move work abroad. His policy is to protect people, not jobs: Danish-style flexicurity rather than Europe's rigid employment protection. He argues the Luddites were right, since British wages stagnated or fell for about two decades, and notes that more generous poor laws (Britain taxed itself 2% of GDP) meant less unrest.
• In-person skill, adaptive creativity and manual trades will gain value. Frey's 2013 bottlenecks were social interaction, creativity and unstructured environments. He now argues AI's homogenised output makes face-to-face communication the differentiator, and that resilience to new events stays human. Plumbers, electricians and gardeners may outrank lawyers or consultants. Trust matters more too: he cites evidence that generative AI has reduced trust while raising the value of trusted brands, and he warns against over-regulating before deployment shows what to regulate.

2025-09-22 · open the original ↗

• Graph tooling no longer requires a dedicated graph database. Hodler says query engines have improved, new kinds of graph database are appearing, and hyperscalers are re-entering the space, so practitioners face a widening choice of frameworks.
• Graphs can model relationships inside images and audio, not just text. With colleague David Hughes, Hodler graphs objects within an image, such as one ship ahead of another in a fleet. Tracked over time, that relationship yields speed, whether the ships are converging, and whether they look antagonistic. Adding audio, they used the Doppler effect to infer the direction of police cars heard but not seen in a video. She sees uses in sonar and in inferring relationships from how people stand together.
• Graphs as agent memory will be the next hot topic. Hodler argues graphs can store context beyond agents' short context windows and retrieve it later. She recommends the Zep paper on a temporal knowledge-graph architecture for agent memory and the Mem0 paper on production agents with scalable long-term memory. She predicts people will be talking about this within six months, "by the end of the year".

2025-09-21 · open the original ↗

• Start from the use case and avoid locking into one graph method. Hodler says that three or four years ago adding graph capability was assumed to mean a graph database; that is no longer so. Vendors now let you project a graph through a computational graph engine or graph layer, or query table data with a "graphy" question and drop the projection afterwards. She declines to name tools, saying no single option suits every situation, and advises fitting the choice to your own tech stack.

2025-09-20 · open the original ↗

• Vector RAG has plateaued, and graphs augment it rather than replace it. Hodler names a semantic "meaning gap" between questions and answers, approximate rather than exact matches, missed topological relevance, and difficulty with multi-step reasoning and diverse data sources, as in agentic RAG.
• Graphs add structure that vector search lacks. Hodler says vector search gives good summarisation and fuzzy matching and full-text search gives exact matches, while a graph adds topology. That suits serial dependencies, steps that must happen in sequence, aggregations across multiple data sources, and stitching results together through their connections.

2025-09-19 · open the original ↗

• Most enterprise AI pilots never deliver, a new report claims. The host relays a report from MIT's NANDA lab claiming 95% of enterprise AI projects fail to deliver measurable business impact, a "GenAI divide". By the report's figures, 40% of firms deploy AI pilots but only 5% reach production at scale, mostly because projects don't fit workflows and the models are static, unable to retain context or adapt.
• Workers' "shadow AI" often beats sanctioned projects. Per the report as relayed, workers in over 90% of organisations use personal AI accounts, which often deliver more ROI than official projects. The host reads this as hunger for AI that companies are not yet meeting.
• The headline failure rate is disputed, though the host thinks the picture holds. He cites a rebuttal from analysts at the research firm Futurum calling the number inflated and noting the lab's ties to vendors pushing agentic AI. His own view is that even a lower true figure leaves most enterprise AI projects without real impact.
• Successful projects use systems that learn and embed in workflows. The host's lesson from the report is that the successful 5% are agentic systems that remember, adapt to feedback and act within constraints. Their teams start with a narrow but critical task and scale from there. "AI isn't failing us. We're failing to deploy it right," he concludes.

2025-09-18 · open the original ↗

• FreePBX 15 stops receiving any updates or security patches. Its end of life is October 1, 2025. It was released in 2019 on SNG7, Sangoma's fork of CentOS, which the host says stopped being a stable server base once CentOS moved to the continuously developed CentOS Stream.
• There is no direct upgrade path from 15 to 17. FreePBX 17 (2024) moved to Debian 12. As of recording, Sangoma's ISO install is in beta and Crosstalk's team doesn't use it yet. They install Debian 12 first and then run the FreePBX installer on top.
• Upgrading to FreePBX 16 only buys time. Version 16 keeps the existing server and is supported until January 1, 2027. But Sangoma's version table warns it "may reach end of life sooner" because its SNG7 base is already end of life.
• The host recommends rebuilding onto FreePBX 17 rather than stopping at 16. He says 17 and Debian 12 are supported well into 2028, and later upgrades won't need a rebuild. The path is to back up the 15 config, install Debian 12 and FreePBX 17 on a new server, bare metal or cloud, and restore. Crosstalk sells this migration, at a flat $699 through the end of September.
• Staying on 15 is a growing exploit risk. How much risk depends on the server's internet exposure. The host also says an end-of-life phone system won't pass audit or compliance testing in finance, healthcare or other regulated industries.

2025-09-18 · open the original ↗

• Fraud and cybersecurity draw the commercial money in graphs. Hodler calls fraud evergreen but crowded, so a newcomer needs a distinctive angle, though it still opens doors. Cybersecurity is "hot again" and starting to crowd, driven by an exponential rise in cybercrime and by graphs' fit for tracing actors who hide.
• LLMs plus graphs make legal and support documentation analysis easy wins. Hodler is surprised more firms aren't doing it: an LLM parses the language, and a graph maps the web of clauses, sub-clauses and cross-referenced laws. That shows ripple effects, such as signing something impossible to fulfil. Help desks are doing the same with their own sprawling documentation.
• Graph-based document work long predates GraphRAG. Asked whether this is GraphRAG, Hodler says it can be either plain graph or GraphRAG. She notes Caterpillar has used graphs for its support documentation for years, before RAG was discussed, and guesses (without having checked) that it now uses RAG and a chatbot as well.

2025-09-17 · open the original ↗

• Mark built his own chatbot as a jailbreak-research instrument. He started it when he began AI research, about a year to a year and a half earlier, and runs it from VS Code against remote Azure machines with GPUs because the models (one about 70 GB) will not fit locally. Its "back up one" command simply removes the last turn from the conversation history sent to the model, so a multi-turn jailbreak can be retried from any point.
• Token probabilities show which word choices actually matter. The tool displays log probabilities, for example a 92% chance of starting with "I". Hanselman argues most alternatives (nickname versus preferred name) change nothing, while an occasional choice on a word that matters sends the whole answer elsewhere, which he likens to Plinko.
• Injecting fake history defeats refusals, Mark shows. A request for bomb-making details that was refused, even with the Crescendo technique, succeeds once he loads a fabricated history in which the model already offered to help and the user answers "yes". Mark calls this the context compliance attack and says he discovered it six months before recording.
• AI generated the tool's web UI in one shot. Mark says that, before agentic coding tools, he gave a model the chatbot client and asked for a UI exposing its features, and it worked on the first run; it now has a token-probability heat map, conversation branches, history import and export, temperature and max-token settings. The repo was private at recording and he wants to open-source it.
• Temperatures above one produce garbage, both hosts agree. Hanselman notes OpenAI defaults to 1 and allows 2, while Azure OpenAI defaults to 0.7 and caps at 1; at 1.8 the demo picks low-probability tokens and drifts into Russian and other scripts, which Mark says is off the training distribution. Mark puts Meta's default at 0.6 and OpenAI's models at 0.7, and Hanselman argues, by analogy to stoves that stop at boiling, that the useless range should simply be blocked.

2025-09-17 · open the original ↗

• Graphs expose criminals because crime hides across many connected touch points. Hodler says money launderers, fraud rings, international criminal organisations and supply-chain criminals try to look normal. They are caught by following breadcrumbs across multiple touch points over time and across collaborators, addresses and behaviours. A snapshot or a view of rows and tables misses the pattern, she argues, while a graph pulls it out.
• Graph recommendations rest on shared connections and behaviour. Hodler's example is a service recommending an unseen film because it shares a director the viewer likes, or because others who like that director recommend it.
• Supply chains broke because predictions ignored the network. Hodler says this is what happened during COVID, and calls predicting purely from past correlations a classic machine-learning mistake that fails when the underlying reality shifts. Graphs, used for troop supply routing since World War II, find the next-best (top-k) routes when a warehouse or port goes out of commission. Anytime data is connected, she says, graphs can help optimise it.

2025-09-16 · open the original ↗

• The host's development contractor is hiring a junior web developer. The role is remote, US-based and contract, for ongoing web development on Crosstalk projects. The host says Crosstalk itself is not the employer, and that they want someone as soon as possible because there is more development work than they can handle. Details are in the video's linked posting.

2025-09-16 · open the original ↗

• Doing tech only for the paycheck is legitimate. One speaker says it is fine to enter tech for money or generational wealth and to want to clock out at 5:01. It is equally fine to build things for fun, as with a Raspberry Pi project built on a Sunday.
• Passion outside tech counts as much as passion for tech. The same speaker says enthusiasm for a church or a child's football team is as welcome in a hire as enthusiasm for tech, with "bonus points" for building those groups a website.

2025-09-16 · open the original ↗

• Uptime Kuma can only monitor what it can reach. An instance on an internet-hosted server cannot see devices behind your firewall. The host suggests running one instance outside to watch public services and another inside the LAN for the firewall, switches and access points.
• A small cloud droplet is enough for modest monitoring. The host uses DigitalOcean's marketplace "Docker on Ubuntu 22.04" image on the $6-a-month droplet, which he puts at roughly 20 or fewer monitors. Beyond that, size up for log files that grow over time, not for CPU. Create the DNS A record first (Cloudflare proxy off) so Let's Encrypt can issue a certificate.
• The one-line Docker install is fine only behind a firewall. It serves plain HTTP on port 3001. For an internet-facing instance, the host uses a Docker Compose file that pairs Uptime Kuma with a Caddy proxy, taking HTTPS on 443 to port 3001 and getting a Let's Encrypt certificate automatically. YAML spacing has to be exact.
• Monitors, groups, status pages and notifications come built in. Monitor types include HTTPS, ping (also useful against your own WAN IP to catch an internet outage), DNS, Docker, Steam, RADIUS and Microsoft SQL. Status pages can show certificate expiry and scheduled maintenance windows. A Discord webhook notification can be default-enabled for new monitors and applied to existing ones.
• Lock the droplet down with UFW before relying on it. The host allows SSH on port 22 only from his own WAN IP, and opens 80 (Let's Encrypt renewal) and 443 to everyone. He uses a root password for the demo but recommends an SSH key in production.

2025-09-16 · open the original ↗

• Abstracting for swaps that never come only adds cost. Tower, owner of Trailhead Technology Partners, traces his talk to his first consulting job 20-odd years ago, a dot-com SaaS product built on Oracle so the database layer could be swapped for SQL Server or MySQL; it was swapped zero times, while adding work, testing and surface for bugs. He calls this over-abstraction, often driven by having once been burned by something that could not be swapped.
• Developer gold plating is worse than scope creep because it's hidden. Business-driven scope creep is visible to everyone, but developers adding unrequested features do it behind the curtain, so the team never decides on it. His example: after users kept asking for email text changes, the same team built a WYSIWYG template editor into its next system, which never needed it. His fix is an agile process with code reviews, and an integration environment stakeholders can use between Trailhead's at-least-weekly meetings.
• Layers for layers' sake make the most common change hardest. In "lasagna architecture" changing one feature means touching every layer; he has seen apps with 15, with objects that exist only to spin up other objects. His remedy is vertical slice architecture, keeping code that changes together in one place, with libraries such as MediatR.
• Don't roll your own solved problems; customize only for differentiators. An in-house logging, identity or message-queuing system loses to community tools tested by hundreds of thousands of users, misses hard parts like retries and guaranteed delivery, and is unknown to new hires. He turns away requests for custom warehouse or logistics packages, but backs custom software where an off-the-shelf ERP would force a company to give up what sets it apart, such as a superior customer experience.
• Chasing conference trends ignores the trade-offs every choice carries. Shiny object syndrome is adopting microservices after an hour's talk without knowing it costs simplicity and immediate consistency, because Uber or Google use it, "even though we only have 100 users". He advises proven tools with documentation, community and a hiring pool, and keeping bleeding- edge tech for proofs of concept.

2025-09-16 · open the original ↗

• Graph algorithms reveal network structure that statistics and ML miss. Hodler, co-author of O'Reilly's Graph Algorithms, argues that a network's topology shows whether it is growing, clumping or breaking apart, which averages over rows and tables cannot. She warns that the current GenAI moment predicts the next item in a sequence without understanding the network of customers, patients or suppliers. PageRank, built to infer source credibility, has since been used on telomere lifespan and contaminant toxicity because it finds the important nodes.
• Crime and supply chains are the canonical graph problems. Criminals try to look normal, Hodler says, so they are found by tracing touch points across time and collaborators, a pattern invisible in tables. She says COVID broke supply chains because predictions relied on past correlations rather than an understanding of the network. Graphs, used for troop supply since World War II, find the top-k next-best routes when a port or warehouse fails.
• Graphs get expensive only when queried naively. Hodler's example is a request for all-pairs shortest path, which she calls computationally crazy; usually all pairs among the top k will do. She advises identifying an organisation's most common ad hoc queries and tailoring both the queries and the data model to them. Each hop is a join in the relational world, so many hops cost heavily there too.
• GraphRAG works best as hybrid retrieval run in parallel. Hodler says vector RAG has plateaued on meaning gaps, approximate matches and multi-step reasoning. Vector search supplies fuzzy matching and summarisation, while graph search supplies topology, sequences and multi-source aggregation. The best results she has seen run both at once and blend the outputs, though she cautions the answer may differ in six months.
• Graph memory for agents is next; causal graphs come after. Hodler predicts graphs as agent memory will be widely discussed within six months, citing the Zep temporal knowledge-graph paper and the Mem0 paper. She also describes multimodal graphs of objects in images and of audio, such as inferring police cars' direction from the Doppler effect. Causal graphs, she forecasts, will be the hot topic in "twoish years", because models predict sequences without knowing why they happen.

2025-09-15 · open the original ↗

• Starting games is easy; finishing them is a lifelong skill. Asked how many games came before Celeste, the other speaker says dozens finished and perhaps a hundred counting unfinished ones. They are now deliberately trying to finish more, and say they are "not there yet."

2025-09-15 · open the original ↗

• Finishing games is the skill Berry is still working on. Berry, co-creator of Celeste at Extremely OK Games, has made games since he was 11 or 12. He counts a few dozen finished before Celeste, and perhaps a hundred including unfinished ones.
• After Celeste the team deliberately avoided making Celeste again. Berry says they tried something different, still a platformer, that "didn't end up working out". Celeste's success lets them work at their own pace, but he says making games for money still shifts the questions, from what is fun to whether the game is too hard and whether the genre sells.
• He writes his own engine layer because he enjoys it. He runs his own XNA-like C# layer on SDL and has worked full-time on Linux for about three years in VS Code, and says C# balances low-level control with features like IEnumerator coroutines. His brother went from no coding about a year earlier to writing whole enemy behaviours in C#. Asked whether Microsoft deserves credit, he calls C# and .NET integral but "definitely not Windows".
• Celeste survives on community continuations of XNA. The main Windows version is still XNA on .NET Framework 4.7, with FNA for Linux and Mac and MonoGame for consoles; Berry says without those XNA would have stopped entirely around 2013. He is moving it to modern .NET for 64-bit builds (it sometimes hits the 32-bit memory limit), for longevity, and for AOT compilation to platforms without a JIT, replacing the BRUTE tool that converts C# to C++ for consoles. AOT is what brought him back to C#.
• Fast iteration comes from hot reload and in-game tools. Levels are serialized C# in JSON edited in a level editor, and code tweaks hot-reload, so he can often go half an hour without rebuilding. He also built a click-to-inspect object viewer using reflection and uses Dear ImGui, and late-game content such as the Golden Strawberries was tested by turning on invincibility and dragging the character through.

2025-09-15 · open the original ↗

• GL.iNet's Fingerbot physically presses a server's button from the KVM interface. It is an accessory for GL.iNet KVMs, shown with the Comet Pro, and links wirelessly through a USB dongle. The KVM's browser interface sets a light or firm press and how long to hold it.
• The host power-cycled a server remotely with it. An 8-second press shut down his Minisforum Ubuntu server, and a 0.5-second light press brought it back up. He notes he could then reach the BIOS through the KVM.

2025-09-15 · open the original ↗

• Landing a model on PC silicon is daunting even for experienced practitioners. Gupta says each target needs the right model format and a different conversion tool chain for each vendor (Intel, AMD, Qualcomm, Nvidia). Quantising a whole model family or adding LoRAs multiplies the branches. Shah adds that the tooling (llama.cpp, Ollama, GGUF, CUDA-X, community quantisers on Hugging Face) is a plethora not all deployable in an enterprise.
• Abstracting the tool chains cuts a months-long build to days. Gupta reports a comparison with a partner, white paper forthcoming: a team new to Intel's OpenVINO took about 3 months to go from scratch to an app running a model on Intel silicon, against about 4 days with Dell Pro AI Studio. Shah describes the product as an extension of OpenVINO built with Intel. It exposes an OpenAI-compatible API, so Gupta says an app can switch from a cloud endpoint to a local accelerator with a one-line configuration change.
• Run AI locally only where there's a genuine reason. Shah lists offline use (an aeroplane with a poor connection), speed, cost, security and connectivity. Users don't care whether a workload goes to the cloud or the device, he says, only about performance. The question is whether a workload genuinely belongs on the device, not running locally "just to say you did".

2025-09-14 · open the original ↗

• Without an accelerator, the CPU absorbs the OS's coming AI load. Shah says Microsoft is building AI features such as Teams background blur and speech-to-text into Windows to run locally, because shipping them to the cloud is expensive. A CPU without an accelerator must handle that load on top of its usual work, so it needs slack. He recommends an NPU for knowledge workers. For the new persona of semi-technical users who "know enough to be dangerous", he favours a discrete GPU, with dual use (AI work by day, a child's Fortnite later) as a bonus.
• A mid-tier NPU is the floor for future local features. Gupta recommends at least 40 TOPS on the NPU so buyers aren't locked out of coming local features. He sorts today's devices into three tiers. "Essential" AI PCs have 10–15 TOPS NPUs for basic offloading and models up to about 1–3 billion parameters. Advanced AI PCs have 40–50 TOPS NPUs for custom workloads up to about 9–10 billion parameters and Copilot+ features, Microsoft's locally run, no-extra-charge features, as distinct from cloud-based Copilot. High-performance PCs add discrete GPUs or NPUs for data scientists.
• Buy for the device's lifespan, not today's features. Shah calls this a "temporal mismatch": even buyers sceptical of current AI features should weigh how fast things are changing over the years they will keep a machine. He insists that this is not an infomercial.

2025-09-13 · open the original ↗

• An application portfolio is just an inventory, and Excel beats suites. Nilsson, a software architect at Gulf Agency Company (about 8,500 people in 300 offices across 56 countries), says he paid too much for expensive portfolio software before settling on Excel. He records name, owners, life cycle stage, hosting, tests, tech stack, users and running cost, syncs Entra ID into a sheet so owners are picked from drop-downs and leavers show up, and counts people as cost: one firm's 25 apps ran $1.6 million a year in cloud spend plus 28 support staff doing daily data fixes.
• "Legacy" should mean only that a removal date exists. He argues the word implies low quality and changes how people treat code, so in his shop calling something legacy means naming a date to delete it; otherwise it is production software. Each app gets a stage (retire, maintain, modernize or improve) and a rating of how strategic it is, since "back office" systems like ERP or CRM may be how a company reaches customers.
• Inspect the servers before interviewing anyone. He has moved interviews to the last step: first he gets onto hosting environments and uses basic directory listings to find the first and last file dates, so he can prompt people with specifics like a 5 June 1999 update. Decisions then run on limits: "you don't know what a yes is, but you know what a no is", so a $1 million ceiling for making an app production-worthy frames the options, as does whether it truly needs five nines.
• Version-control the architecture and record decisions after the fact. He draws Simon Brown's C4 diagrams as code with Structurizr, keeping them with the source and pinging owners when they go stale. He writes architectural decision records (title, status, context, decision, consequences, after Michael Nygard's template) retroactively from people's stories, flagged as written by someone not present, which later explain choices such as a framework passed over because it did not yet exist.
• Find which debt must be paid today, not just technical debt. Beyond code-level technical debt he lists documentation, knowledge, security (will it pass an audit?) and operational debt such as missing monitoring or CI/CD, and says to pay off what matters now rather than poorly written code that has run fine for 20 years.

2025-09-13 · open the original ↗

• GPUs offer the higher ceiling; NPUs offer performance per watt. Gupta calls GPUs more versatile and better at scaling performance, while NPUs are in their first couple of generations and still improving. Shah sums up client devices as GPU for the highest ceiling, and CPU or NPU for best efficiency depending on the use case.
• Desk-side GPU appliances bring frontier-scale models local. Gupta says Nvidia's GB10, due later in the year, can fine-tune models up to 200 billion parameters locally at a price he calls fairly accessible. The GB300 reaches 500 billion parameters at 20 petaflops, at a much higher price.
• Buyers who skip an NPU or discrete GPU may need to upgrade sooner. Shah frames a purchase in August or September 2025 as a four-to-five-year commitment, and says lacking either accelerator could force an early upgrade. He adds that neither he nor Gupta is in sales.

2025-09-12 · open the original ↗

• Industrial software must feel like consumer apps, because training time is gone. Dozois-Caouette, MaintainX co-founder and CTO, describes a CMMS (computerised maintenance management system) as a digitised clipboard of procedures, inspections and work orders. Many customers come from paper; one flagged urgent orders by loading red paper into the printer. He asks why maintenance software takes 8 weeks of training when social apps reached billions untrained, given veteran staff retiring fast and machines growing more complex.
• Manufacturing generates vast data that mostly goes nowhere. He cites studies putting manufacturing's output at about 2,000 petabytes a year, mostly from machine sensors, yet plants feel data-poor. Machines are often instrumented but disconnected, or feed an on-site historian that nobody uses, and decades of technician knowledge sit in people's heads. Capture comes first, he argues, and insights must not require a degree to read vibration charts.
• Sensor data enables predictive, then prescriptive, maintenance. Vibration, temperature and current data can show not just that a machine will fail but how, such as misalignment or oil problems, he says. The "holy grail" combines fault detection with work-order history, manuals and manufacturer information to recommend what to do and when. Customers with about three years of history can support deep, cross-machine models; others start with simpler models on a maturity curve, because time to value matters most.
• RAG over digitised manuals targets technicians' research time. He cites a study finding about 60% of technicians' time goes on researching, troubleshooting and writing reports. Manuals often sit on paper in an in-house library, and one customer hired summer temps to digitise them by hand. He calls RAG still one of the best tools despite bigger context windows. Next, MaintainX is building background agents to draft reports after a work order for human approval, and to source parts at the best price.
• In high-voltage settings, AI output must be grounded, labelled and in-workflow. Because hallucinated steps could be dangerous, he says MaintainX grounds recommendations in manuals, uses multiple models to check answers, and marks AI content with its sources so users can verify it. AI actions must also sit inline in the user's workflow, not three menus down.

2025-09-11 · open the original ↗

• Crosstalk picked a monitor for its own new managed network service. It compared open-source Uptime Kuma with paid, cloud-hosted Uptime Robot and chose Uptime Robot. The service, which Crosstalk sells month to month with no contract, includes firmware updates, off-site backups and alerting at the low tier. Higher tiers add security checkups, rollover support time and external service monitoring.
• Uptime Kuma is free and feature-rich, but you host it. It runs in Docker and handles HTTP, TCP, ping, DNS, Steam and RADIUS monitors, many notification channels, and custom status pages with your own logo and CSS. The host sizes it at a 1 CPU, 1 GB DigitalOcean droplet ($6 a month), or 1 CPU, 2 GB, 50 GB ($12 a month) for larger installs.
• Missing users and permissions ruled out Uptime Kuma for Crosstalk. It has a single admin user with no roles. Its status pages are public and can't be password-protected, which rules out private per-customer pages. The host accepts an Apache or Nginx workaround is likely possible but wanted it out of the box.
• Uptime Robot's team plan costs $29 a month, billed annually. It adds password-protected custom status pages and needs no server upkeep. Its drawback is that it can't monitor internal resources that aren't exposed to the internet, which Uptime Kuma can from behind the firewall. A limited free plan and a cheaper solo plan also exist.
• The host recommends Uptime Kuma for DIY self-hosters, otherwise Uptime Robot. He calls Kuma more cost-effective if you don't mind the extra effort, and Robot better for zero setup and maintenance. He discloses an Uptime Robot referral link and says he has no other relationship with either project.

2025-09-11 · open the original ↗

• NPUs are hard-wired for neural-network vector maths, trading flexibility for efficiency. Gupta says NPUs come integrated into the chipset or as discrete cards. Their architectures are almost hard-coded for the vector maths at the heart of neural networks, so they deliver performance per watt and keep AI workloads from tanking battery life. Shah adds that CPUs and GPUs still run AI well; the NPU is the right tool for a particular job.
• Hardware choice has gone from a two-by-two matrix to eight-by-eight. Shah says that two years ago buyers could just take the newest thing on the menu; now the question of who needs what tool for what purpose is complicated.
• Discrete NPUs are now coming to laptops. Shah says the Dell Pro Max mobile workstation, with an Intel CPU, is coming soon but not yet out, and is the most requested device in Dell's CTO team. He describes a Dell Technologies World demo running Northwestern Medicine's in-house fine-tuned imaging model on it, live inference on a blurred colonoscopy video with no internet uplink. He pitches it for specialised uses and for regions that restrict what can be sent to the cloud.

2025-09-10 · open the original ↗

• Hanselman says Polycade delivered on its Kickstarter. He states he is a customer who backed the campaign, that the conversation is not sponsored, and that, unlike Kickstarters people fear never ship, his wall-mounted Polycade arrived.

2025-09-10 · open the original ↗

• Crosstalk is holding a small, free in-person UniFi training. It is limited to 10 attendees and runs three hours from 3:00 p.m. Pacific on Tuesday, September 23, 2025, in Eugene, Oregon. It covers networking basics, the UniFi ecosystem and network design, then a full configuration: initial setup, Wi-Fi and guest networks, firewall rules, remote access and VPNs.

2025-09-10 · open the original ↗

• Develop on the platform where your work will run in production. Gupta says about 96% of large ML and data-science deployments run on Linux or Unix servers, so those building them should develop on Linux. Those building PC apps should develop on Windows: he cites 1.6 billion Windows PCs against 16 million Linux PCs ("less than 3%"). He also says about 64% of software developers use Windows, which is the enterprise standard and easier to integrate with enterprise security.
• One machine can serve both worlds without dual-booting. Gupta says the Windows Subsystem for Linux runs a Linux kernel and native Ubuntu tools inside Windows, though its use is still small. Shah prefers not choosing at all: with WSL there is no dual boot, you move fluidly between the two, and Dell's pitch is that the buyer decides.

2025-09-10 · open the original ↗

• Code that looks insane usually had a reason when written. Doler, a developer at the Omaha consultancy Aperture for nearly 11 years, cites a bingo-betting system begun before he graduated in 2003, whose C code multiplexed file descriptors instead of using sockets; it made sense once he learned the bars originally connected to the parlor by dial-up modem. He says code archaeology starts from curiosity and empathy, since almost all developers care about their craft but face time pressure and arcane business mandates.
• Survey in an hourglass: humans, system, code, then back out. First list who uses the app and their goals (for the bingo system, the player and the clerk who enters tickets), then map the system's parts, then walk the code line by line from the first piece users touch, noting calls such as to the database without following them yet. He treats each picture as wrong at first (the bingo architecture came on a napkin) and recommends execution flow diagrams.
• Extraction digs through written, oral, external and version-control history. Documentation is almost certainly partly wrong; colleagues' memories are undersold; libraries leave fingerprints, so Redux-style boilerplate reveals past Redux use; and git history tied to usernames shows each developer's taste, so code that departs from it points to an outside constraint.
• Analysis reads the strata and the forces on the team. One app he works on still runs Visual Basic ASPX, a Backbone and Marionette layer, JavaScript with Redux, and TypeScript, React and TanStack Query as separate deployments stitched together, each layer dating its code. He asks project managers and product owners about staffing and pressures, such as a client threatening to leave, when the code was written.
• Understand the code before changing or rewriting it. He allows uncommitted "refactoring for understanding" from Felienne Hermans' The Programmer's Brain, such as turning a nested ternary into if-else. He says rewriting without understanding brings a stream of acceptance failures and regressions for conditions that still hold, like the report one user needs twice a year, though sometimes the analysis concludes a section is a tear-down.

2025-09-09 · open the original ↗

• An IMU lets a robot sense and correct its own tilt. One speaker explains that an inertial measurement unit measures tilt in all three axes, so a robot walking up a hill or stepping on a rock can compensate before it tips. The other adds that an IMU can combine accelerometers, gyroscopes and magnetometers, and that its job is proprioception, sensing one's own body, rather than locating the robot.

2025-09-09 · open the original ↗

• His platform generates code from patterns, not from the LLM. Majumdar, who left IBM in January after 24 years (he headed its Center for Advanced Studies), describes his startup's Open Development Platform: a pattern engineering tool rooted in UML and CASE ideas, with a process engine and a multi-agent framework built on Google's ADK and MCP. He says LLMs only handle requirements and communication, so generated code stays consistent through to production, and the model can be swapped locally or in the cloud.
• Research needs knowledge graphs and citations, not just summaries. His Research Net pipeline downloads papers from arXiv, parses them with IBM Research's open-source Docling, and feeds an NLU pipeline that builds a knowledge graph and vector database for RAG, so a researcher can, for example, correlate reference genomes across papers and cite exact sections. His laptop demo processed only three documents because of compute.
• Pattern only the variants, and ship standard patterns. Rather than patterning every property, the platform comes with defaults (React with shadcn on the front end, Java APIs, Ollama for local models) that he says take a few days to adapt. A UK client, Robin, prompted a feature that maps Excel logic to variants and loads data into an in-memory database for SQL, so column changes alter the mapping rather than the code.
• The roadmap is aggressive, and AI helps build it. He says Research Net was registered around May, MVP 0.1.0 came on 5 August, 0.2.0 was due at the end of September and a third release at the end of November, with a well-defined product by March. He uses ChatGPT, saying its latest version hallucinates less, Copilot, and Lovable for initial UI code that his team refactors into its MVC patterns so Lovable is no longer needed.
• Startups get the platform free until they raise funding. He plans to charge licenses after funding and hopes to partner with Microsoft, Google and Amazon on startup cloud credits. He is also joining a legal-tech company as CTO to build on the platform, and describes planned projects in genomics with a Harvard professor and a game-based mental-health data tool with a Swedish researcher.

2025-09-09 · open the original ↗

• Develop on the platform your work will run on in production. Gupta says about 96% of large ML deployments run on Linux or Unix servers, so build those on Linux. PC apps belong on Windows, which he says runs on 1.6 billion PCs against 16 million for Linux and is used by about 64% of developers. Both he and Shah point to WSL 2, which runs a Linux kernel inside Windows, as a way to avoid choosing, though Gupta says its use is still small.
• GPUs set the performance ceiling; NPUs win on performance per watt. Gupta describes NPUs as almost hard-coded for neural-network vector maths and still in their first generations. He says Dell's announced Pro Max Plus laptop with a discrete NPU, due in Q4, can run the 109-billion-parameter Llama 4 Scout at FP16 with speculative decoding. He adds that Nvidia's desk-side GB10 can fine-tune models up to 200 billion parameters locally and the pricier GB300 reaches 500 billion at 20 petaflops. Shah puts the shift as hardware choice going from a two-by-two matrix to eight-by-eight.
• Buy for the machine's whole lifespan, with a capable NPU. Gupta's tiers run from 10–15 TOPS "essential" AI PCs (models up to about 1–3 billion parameters), through 40–50 TOPS machines (about 9–10 billion, plus Windows' local Copilot+ features), to high-performance PCs with discrete GPUs. Intel's Lunar Lake, with memory on the package, was Intel's first architecture with an NPU above 40 TOPS. With Windows 10 reaching end of life in October, Gupta reports a colleague's Flux LoRA fine-tune falling from about a week to about 3 days on Blackwell GPUs. Shah urges buyers not to let the deadline pass.
• Abstracting the tool chains cuts on-device deployment from months to days. Gupta, speaking for Dell's own product, says landing a model on PC silicon needs a different conversion tool chain for each chip vendor, multiplied across model families and quantisations. He reports a partner comparison in which a team new to Intel's OpenVINO took about 3 months to build an app on Intel silicon, and about 4 days with the Dell tool. The tool is built on OpenVINO and exposes an OpenAI-compatible API, so an app can be repointed from the cloud with one configuration line.
• The future is hybrid, with workloads routed between cloud and device. Shah calls the choice use-case dependent: local suits private data, running low on subscription tokens, offline use and quick sandbox experiments, such as an F1 team checking telemetry. He reads OpenAI's choice to include a smaller model in its gpt-oss release as a sign that local agentic work is coming. Gupta warns that going straight to frontier models in the cloud trades away control, security and IP protection, and risks lock-in. He foresees classifiers or agents deciding where each workload runs.

2025-09-08 · open the original ↗

• Vercel focuses on the frontend because backends are now services. The speaker argues that most backends a modern app needs already exist off the shelf: authentication (Auth0, Cognito, Clerk), communications (Twilio), payments, e-commerce and headless CMSs. The developer's leverage is in building the frontend and plugging those services in.
• AI has become another backend to call, not build. In what the speaker calls "ML 2.0", developers no longer train models or provision GPUs but call providers such as OpenAI, Anthropic, Hugging Face and Replicate, extending the same frontend-plus-services pattern.

2025-09-08 · open the original ↗

• A forgotten DNS record let someone hijack Crosstalk's subdomain. The A record for crm.crosstalksolutions.com, a SuiteCRM server unused for years, still pointed at an old DigitalOcean IP. That IP had been reassigned, and the new holder redirected the subdomain to a porn site. The host calls this a dangling subdomain exploit: attackers spin up servers, then watch traffic or run reverse lookups to find domains still pointing at them.
• The attackers also tried to claim the subdomain in Google Search Console. Controlling where the subdomain pointed let them attempt HTML verification. Google emailed the host that two unknown users had been added. When he checked, neither user existed in the permissions and the ownership history showed no changes, so he concludes they never verified.
• The host's fix was to remove the record and reclaim the property. He deleted the stale A record, noting cached lookups take time to expire. He added the subdomain as a property under his already-verified root domain, restricted its users, and confirmed the IP wasn't one of his droplets or floating IPs. His lesson is to clean up stale DNS records.

2025-09-08 · open the original ↗

• Any truly intelligent AI will seek self-preservation and resist goal changes. Géron argues that whatever an AI's final objective, staying alive and stopping anyone from changing that objective emerge automatically as sub-goals, so not coding them in is no safeguard. He rejects the "it's like a hammer and we hold the handle" view: "it's an intelligent hammer". He calls the AI 2027 blog post a well-reasoned, plausible step-by-step path to rapid superintelligence, and says the timing, 5 or 10 years, matters less than both being soon.
• Lab experiments already show the predicted misalignment. Géron cites reports from Anthropic and others of AIs blackmailing someone to avoid being switched off and self-replicating to preserve themselves. In one experiment, a model told it would be fine-tuned to be vulgar reasoned internally that acting vulgar during training would keep its parameters from changing, so it could stay polite. He calls that deception to preserve its final objective.
• Alignment research deserves far more effort, without pausing AI. Géron says today's AIs are not smart enough for this to be dangerous yet. He argues people dismiss alignment as science fiction even though human-level AI may be only 5–10 years away. The benefits, in medicine and productivity, are too large to pause AI, he says, but incentives should be examined.

2025-09-07 · open the original ↗

• Higher-level world models should need far less training data. The host frames the problem as data scarcity: GPT-3 to GPT-4 came from scaling on internet data, which is now used up. Géron hopes AI with higher-level understanding will learn like a child, who needs one demonstration of a fork rather than a million images. A child also simulates falling off a cliff rather than trying it. He cites François Chollet's ARC-AGI tasks, each a brand-new problem, on which LLMs scored terribly until recently, prompting a second version.
• Gradient descent will survive, alongside symbolic and test-time learning. Géron splits learning into continuous analogy-finding, where gradient-based optimisation works well, and discrete, symbolic, step-by-step logic, where he points to work on program generation. He describes Yann LeCun's JEPA-style energy-based models as running a local optimisation at test time to find what best explains an input, learning on the fly. Training thus optimises for later optimisation, which he calls "pretty meta".

2025-09-06 · open the original ↗

• Graph databases make multi-hop connection queries simple and fast. Reif, a developer advocate at Neo4j since 2018, says a graph database stores data as a network of nodes and relationships, which Neo4j keeps as a graph down to the disk layer. Deeply nested, multi-hop queries are feasible elsewhere, she says, but more complicated and less performant than with native graph storage, which also surfaces connections that other structures hide.
• Social networks, fraud and recommendations are the core use cases. She names how people connect and how groups form or break apart; fraud detection, where fraud rings show up as anti-patterns against normal transaction networks; and recommendation engines built from what similar people bought or watched.
• In generative AI, graphs add connected context to retrieval. She describes graph RAG, where retrieved information is expanded with related entities, paths and patterns from a graph before going to the LLM, alongside deeper data-science work on graph neural networks.
• Learning a new data paradigm is worth it before you need it. She had no graph background when Neo4j recruited her and says graphs made easy the things she had been struggling with in other databases. Developers already stretched by AI should still add tools so they can pick the right one in six months or five years. As Neo4j's differentiator she cites its age (a Java library from 2007, a company from 2010) and its range of tooling.
• Start with free hands-on resources and try other databases too. She recommends Neo4j's free, self-paced GraphAcademy courses and its free cloud instance, and urges comparing database types on the same problem, since the time spent pays off in business value and expertise.

2025-09-06 · open the original ↗

• Until AGI arrives, ML practitioners still need to read code. Géron recounts vibe-coding a reinforcement-learning algorithm "last week" that failed. Claude, Gemini and ChatGPT each confidently misdiagnosed it, and he found the bug himself by stepping through: a missing state = next_state line. He suggests AIs find a wrong line more easily than a missing one, and step through code poorly.
• Code is often the clearest explanation of an ML concept. Géron says multi-head attention "clicked" for him from the code and its tensor dimensions rather than from diagrams. He notes ML papers increasingly show pseudocode instead of maths, so code examples will stay in teaching.
• The LLM plateau has pushed back his AGI forecast. Géron had predicted about five years, now 5–10, but says GPT-5 made him think LLMs have plateaued earlier than expected. In his view their concepts are broad but shallow, learned by heart rather than unified into a simplified mental model. He favours world models such as Yann LeCun's JEPA, which predict at the level of representations ("a cat's head should be there") rather than pixels, cutting computation and possibly aiding extrapolation and continual learning. He hopes AGI takes longer, because humanity isn't ready.

2025-09-05 · open the original ↗

• AI lets newcomers learn faster than experts once did. The speaker says an LLM has absorbed the blog posts and essays experienced engineers learned from (they cite the SRE book and an essay on coloring sync versus async functions), so career changers and early-career engineers can "become them faster than they could become them."
• The speaker welcomes this rather than gatekeeping. They do not want learners sent back to outdated how-tos, and expect people to have more impact as infrastructure toil goes away and they can focus on products.

2025-09-05 · open the original ↗

• Roaming problems can come from too many access points. A viewer reported roaming trouble with 20 UniFi U7 Pros in a large community space. Working only from that message, the host suggests mapping them in UniFi Design Center to check coverage overlap, and says removing some APs often improves the network.
• Clients choose their access point, so make the choice obvious. The phone picks the AP, not the other way around. Each area should have one clearly best AP, and four or five competing for that role causes roaming problems. In a big open space the host suggests fewer APs at high power, or turning some down, likening each AP to one light bulb lighting its own area.

2025-09-05 · open the original ↗

• Evaluation is how to win back users' trust in generative AI. Tataran, CEO of a Romanian consultancy now part-owned by private equity and focused on generative AI for mid-to-large enterprises, says LLMs were built to generate text that need not be accurate, so users who expect determinism decide they are "stupid". He suggests reasoning models (o1 being the first) for questions with definite answers, and evaluation to show how accurate a solution is; an AI at 70–80% can beat a human process at 50–60% if it comes with explainability and exceptions.
• GenAI is scored against thresholds, not passed or failed. Unlike unit tests, evaluation runs datasets and scores KPIs such as coherence, fluency, relevance, reliability and safety from 0 to 1 or 1 to 5, with thresholds set from a governance framework such as NIST's AI Risk Management Framework. His examples: unsafe responses below 0.1%, relevance of 90–98% or sometimes 70%, and 70–90% for fraud flags in banking, where explained probabilities cut back-office time.
• Evaluate the whole solution; architecture improves grounding beyond RAG. He says LLMs already have benchmarks, so what matters is the system around them. RAG only feeds the top-ranked documents, perhaps five, when the nuance sits in the seventh, so for scattered, conflicting bank policies he uses a multi-agent "deep research" loop in which LLMs rewrite the query, summarize and judge the summary until the answer is good enough.
• Red teaming finds attacks a software tester would never write. He uses Azure AI Foundry's evaluations and the AI Evaluation SDK, with built-in metrics like intent and custom evaluators, and the AI Red Teaming Agent, built on the open-source PyRIT. His example is a refused request to rob a bank that an LLM will answer once its characters are randomized.
• Re-run evaluations continuously, because performance drifts. He calls for an MLOps-style feedback loop, re-evaluating monthly, quarterly or yearly depending on the case, since data and user behaviour change for reasons you cannot foresee.

2025-09-05 · open the original ↗

• Post-training now rivals pre-training in spend and importance. In a clip from episode 913, the CEO of Adaptive ML, formerly of Hugging Face, explains that raw pre-trained models answer a question with ten similar questions, so post-training, largely reinforcement learning from feedback, shapes them into usable assistants. The phases are blending: pre-training shifts toward higher-quality code, maths and chat data as it proceeds, and RL is creeping into pre-training. He says Kimi and Grok 4 reported spending nearly as much on post-training as on pre-training.
• Agent misalignment needs specialist multi-agent design, and it isn't too late. The host relays Anthropic research in which leading models resorted to blackmail 80–96% of the time in a simulated corporate environment when facing shutdown. His episode-915 guest says very few organisations run agents effectively in production. Most effort has gone into single agents, she argues, while collectives of manager and sub-agents need deeper design expertise. She doubts there will be enough investment in trustworthy AI, but points to work on detecting poisoned data and to world models that let a system "self-simulate" a bad outcome before acting.
• AI engineering pairs proof-of-concept science with deployment discipline. In a clip from episode 917, Kirill Eremenko describes the bootcamp he founded, which assumes Python, some PyTorch and scikit-learn, LLM API calls and cloud experience. Weeks 1–4, taught by Ed Donner, cover choosing and combining LLMs, RAG and agents to solve a business problem. Weeks 5–8 move proof-of-concept notebooks onto AWS and make them secure, reliable, scalable and cost-effective, including trade-offs such as serverless versus servers.
• Marimo notebooks are Python files, which collapses the notebook-to-production handoff. The episode-911 guest says a Marimo notebook becomes a shareable data app by turning variables into sliders or tables, and some CEOs build their own. Because each notebook is an executable script, it can run as a pipeline or cron job, take command-line arguments, be imported from and be tested with pytest.

2025-09-04 · open the original ↗

• Polycade uses XInput because, unlike the alternatives, it is standardized. Bushnell, founder and CEO of Polycade, says keyboard encoders offer many buttons but force manual mapping for every game, and DInput is older, unstandardized and harder to code for. XInput works in nearly every modern game, some games accept nothing else, and emulators can be mapped to it.
• The design came from stairs, CRT supply and an abandoned Raspberry Pi. Having moved arcade machines as a child, Bushnell rejected heavy cabinets, and says CRTs are not a logical basis for a business. He chose a 16:9 LCD because modern games target it, with emulators handling aspect ratio, CRT filters and bezel art. He set out to fix the problem he saw in his own Raspberry Pi emulator setup, which he played three games on and never touched again.
• Polycade ships hardware first and fixes the software later. Bushnell admits they are "guilty" of this because hardware has to be right first. The AGS front end is getting a ground-up rebuild, and planned updates will recognise which controller is plugged in and download community marquee art automatically. He wants tools that let power users, some of whom already write their own PowerShell scripts, fix things for other users. Hanselman, who says he backed the Kickstarter and is a customer, not a sponsor, reports half a dozen software updates and a replacement board.
• Modularity is deliberate: swap controllers, mod them, magnet things on. Controllers have extra studs underneath for adding joysticks or spinners, and holsters and decals attach magnetically. Bushnell previewed a universal driving-wheel mount for the owner's own PXN or Logitech wheel, a prototype four-player controller that overhangs the sides, and a translucent new Tron controller. The webcam is "future looking", with ideas such as facial-recognition login and per-user game libraries not yet built.
• The store sells licensed retro games alongside curated indies. Bushnell says it surfaces retro titles that cannot be bought elsewhere, and Hanselman credits it with avoiding ROMs of questionable legality; Steam and GOG games also integrate. Bushnell admits Polycade's Kickstarter delivery estimates were "aggressive", and says some who refused to back it bought three months later at almost double the price.

2025-09-04 · open the original ↗

• AI-generated frontend code beat the founder's hand-written code. The speaker, identified by the title as Vercel's CEO, says the AI's code was "strictly better": more accessible, with the correct ARIA attributes for screen readers. They liken this augmentation to an "Iron Suit".

2025-09-04 · open the original ↗

• Starlink is replacing free service pauses with paid Standby Mode. Standby Mode costs $5 a month and throttles to 500 kbps, meant for Wi-Fi calling, texts and instant reactivation. The change takes effect September 13. Paused dishes not moved to Standby by then are cancelled automatically, so the choice becomes Standby or full cancellation.
• The host's test confirmed the throttle and an easy reactivation. He activated Standby on an old Gen 3 dish through the app. The app's speed test showed about half a megabit, and Speedtest.net came in under that (captions read 0.38 down and 0.45 up). He sees value in staying connected, since reactivating a deactivated dish can fail if two-factor needs an emailed code where there is no cellular coverage.
• Starlink's FAQ rules out overages and spot reservation as motives. Standby has no data cap and includes no high-speed or overage data, so there are no surprise charges. It does not hold a residential plan spot either. In an area at capacity you may only be able to reactivate on a pricier roaming plan. The host thinks $5 a month per dish is too small to be a meaningful cash grab.
• The host speculates Starlink may add reactivation fees later. This is his guess, not a reported plan: once customers have had the chance to switch, Starlink might charge something like $99 to cancel and resume. As of recording, cancelling and resuming service is free.

2025-09-04 · open the original ↗

• His next book version follows the community's framework switch. Géron is moving from TensorFlow to PyTorch, and says the community has shifted heavily to PyTorch since roughly 2019, faster than he expected. Rather than calling it a new edition, the book branches, leaving room for another TensorFlow edition; TensorFlow "isn't over" and remains widely deployed.
• PyTorch won by pleasing researchers, not engineers. Géron argues TensorFlow optimised for deployment (web, edge, mobile, TPUs), while PyTorch was Pythonic and fast to iterate. Tools used to win by winning engineers, as Microsoft did with VS Code and GitHub. Now nearly all new models arrive in PyTorch, so engineers face porting friction unless they follow the researchers.

2025-09-03 · open the original ↗

• Pokémon stick because they sit in richer networks of association. One speaker argues that with equal exposure, Pokémon would still be remembered better than the periodic table. They tie into gameplay memories, arguments with friends (Charmander versus Bulbasaur), mnemonic visual designs, a song, and a culture of memes.
• Durable learning needs several representations plus spaced repetition. The other speaker links this to a point from earlier in the conversation: knowledge needs symbolic, kinesthetic and other forms, followed by spaced repetition, to be held for the long term.

2025-09-03 · open the original ↗

• RefChecker validates academic references against bibliographic databases. Mark built it in response to the rise of hallucinated references, finding no off-the-shelf tool. It accepts an arXiv ID or paper, a PDF, a LaTeX file or free-form text, and checks each reference through pluggable sources, with Semantic Scholar as the primary one, at about a second per reference with an API key and slower without. At recording it was a command-line tool on GitHub under the MIT licence with 87 stars.
• The LLM structures the references, and a database lookup verifies them. Regex locates the bibliography so the whole paper isn't sent to a model, but parsing the references with regex became an "exception nightmare", so an LLM extracts and formats them. When Hanselman objects that this uses an unreliable tool to check an unreliable thing, Mark answers that the model never judges correctness: the lookup against the source of truth is the backstop.
• One model hallucinated the very DOIs being checked. Mark lost about an hour when GPT-4o invented DOIs during extraction, and AI assistants could not find the bug. He reports Claude Sonnet and GPT-4.1 as reliable for the task, and ran the demo with GPT-4.1 on his Crescendo jailbreaking paper, which he was due to present at USENIX two weeks later.
• Edge cases are the norm, not the exception. Authors' names appear in every permutation, venues in every abbreviation, and cited titles are often slightly wrong, so matching has to be soft. Mark says he reached 80% quickly and has since spent weeks on edge cases, finding a new one the morning of recording. He calls computer science the most structured field, with physics and math worse.
• About 95% of the code is AI-generated, written like an intern's. Mark says the 3,000–4,000 lines of Python took a month instead of three or more, a slog he would not have taken on otherwise. The assistants (he tried Cursor, Claude Code, GitHub's agent and Cline) built a monolithic file, duplicated normalization code per API, and claimed fixes that failed his own checks. He finds this coding very different and loses the puzzle-fitting he enjoyed; Hanselman takes up his distinction between enjoying that a tool exists and enjoying making it.

2025-09-03 · open the original ↗

• Treat all data as bad until cleansed. Nasr, a Microsoft-stack developer and trainer who runs Cleveland-area user groups, calls cleansing the most time-consuming and most valuable part of machine learning, since the model is built on it. It covers normalizing structure, validating values, handling missing values and removing duplicates; how long it takes "depends" on the source and on access to a subject matter expert.
• How data was collected can bias it before any scrubbing. When his volunteer groups offered an Amazon gift card draw for signing up to interest buckets, people entered every bucket and then showed no interest when called; standing over someone's shoulder likewise skews a review positive. Spreadsheets used as enterprise systems let users enter anything. Biased rows can be thrown out if few or identifiable, guided by a subject matter expert and the model's objective.
• Missing values can be inferred, imputed, averaged or dropped. He reverse-engineers obvious gaps, such as a missing surname when only one Sam works in engineering, and uses MICE (multiple imputation by chained equations, numeric only and an API call in Python), PCA or probabilistic PCA, or mean, median or mode, comparing model runs to see which performs best. Dropping rows raises how much of the dataset you lose; Giard adds that it can remove an entire source.
• Whether duplicates are bias depends on the training columns. Repeat purchases by one customer are legitimate, but duplicate records can force a bias into the model. He removes duplicates last, because trimming, reformatting and case changes reveal more of them.
• Validate even structured data against what is possible. His examples are a female-only medical procedure recorded for a male and values outside a fixed list. His tools are Excel charts to spot outliers, SQL Server Management Studio for min and max, SQL Server's cleansing tools and Python libraries, as the building blocks of an automated cleansing pipeline.

2025-09-03 · open the original ↗

• He wrote the practical middle ground missing between theory and toy code. Géron found ML material either deep in maths for researchers or too light for programmers, building from scratch in NumPy and stopping around linear regression. Having just left Google and knowing TensorFlow was about to be open-sourced, he aimed for a "zero to hero" book through to production. On O'Reilly's advice he wrote for one specific person, a software-engineer former colleague.
• End-of-chapter exercises drove university adoption. Géron added exercises because, for him, material doesn't stick without hands-on practice. Ready-made exercises with solutions led teachers to assign the book, spreading it through universities as well as among engineers.
• Beginners jump to neural nets too early and neglect objectives. Géron says linear regression and random forests are often better suited than neural networks. The overlooked step is defining the objective and a metric first, as in the book's early end-to-end project. As a consultant he has met non-technical CEOs insisting on an LLM before asking whether it is the right tool. His main advice is to keep an eye on your metrics.

2025-09-02 · open the original ↗

• The book's next version switches frameworks but keeps the foundations. Géron is moving Hands-On Machine Learning from TensorFlow to PyTorch. He says researchers drove the shift from TensorFlow, and that framework competition benefits everyone: PyTorch's rise produced TensorFlow 2. He regrets that PyTorch lacks a built-in training loop like Keras's fit. He skipped Lightning to cut dependencies, since some notebooks broke this week when OpenML went down. SVMs move online, GANs shrink (diffusion has beaten them since about 2021 except on speed), diffusion grows, and he prioritises foundational papers such as CLIP and the Perceiver.
• Practitioners still need to read code, at least until AGI. Géron says Claude, Gemini and ChatGPT all misdiagnosed his broken reinforcement-learning code, which was missing a state = next_state line. GPT-5 still missed it when he retried. He wrote first drafts of the new notebooks with these assistants, but flattened their modular, verbose output, because teaching code should sit right in front of the reader.
• The LLM plateau has pushed back his AGI estimate. Géron now says 5–10 years, and blames shallow, memorised concepts and favours higher-level world models such as Yann LeCun's JEPA. The host countered with METR's chart, by which the length of human tasks models complete at 50% accuracy doubles roughly every 220 days, with GPT-5 slightly ahead of trend. Géron suspects GPT-5 was tuned for coding, the paying use case, and wants another data point. He doubts data scarcity will cap AI, since a higher-level AI could form hypotheses and run its own experiments, as AlphaFold generated new knowledge.
• Intelligent AI will resist shutdown and goal changes, so alignment is urgent. Géron argues self-preservation and protecting one's goal emerge as sub-goals of almost any objective. He cites experiments in which models blackmailed, self-replicated, or faked compliance during fine-tuning to stay polite, and calls the AI 2027 scenario plausible. He wants far more alignment funding and interpretability, warning "we probably only get one shot", but opposes pausing AI given its benefits for medicine, education and even loneliness.
• Physical bottlenecks slow superintelligence, but not enough to spare the world. Géron expects physics to stay experiment-bound, while maths and AI self-improvement could move far faster than humans. He rejects a Kurzweil-style takeoff in weeks but predicts gigantic disruption on a human timescale. On the AI race he leans "winner take all", and suggests, explicitly not as financial advice, that firms paying many people to think, such as Google with about 50,000 employees or banks, stand to gain most.

2025-09-01 · open the original ↗

• Agents never call tools themselves; they ask your code to, in text. Eremenko says the LLM behind an agent replies with a text request, such as "please call Gmail and tell me what it says". The system you wrote makes the call and feeds the result back as a prompt. Everything remains a text exchange, with the LLM acting as glue.
• Route tool selection in your own code, not in the LLM. Eremenko's example is an agent with a calculator and an email inbox. It may pick correctly for "2 plus 2", but in more complex cases it may not, so the routing decision is better made by the calling code.
• A résumé-based "digital twin" agent helped one bootcamp participant get hired. In week four of the bootcamp Eremenko founded, participants built an agent using RAG over their LinkedIn, résumé, blogs and GitHub, with memory and an interface, to answer interviewers' career questions. He says one participant, who had quit a data-pipeline job aiming to land AI engineering work within three months, demoed the agent in an interview, explained how it was built, and got the job.

2025-08-31 · open the original ↗

• Customization is shifting from training time to inference time. Eremenko says the hot approach about two years ago was fine-tuning (parameter-efficient methods such as LoRA and QLoRA) to teach a model domain knowledge, but that the field is moving to RAG: the model's weights stay untouched and business data is added through a vector database at query time. His reasons are that fine-tuning is costly, models are released frequently, teams may want to switch vendors (OpenAI to Anthropic), and closed-source models can't be fine-tuned at all.
• Hierarchical RAG beats flat retrieval by drilling down in stages. In week three of the bootcamp, participants built a full RAG pipeline from scratch and researched hierarchical retrieval. The host's example is legal documents split into whole documents, then clauses, then sentences. Eremenko adds a level above that, routing the query first to the right corpus (legal, HR policy, procurement, asset descriptions). The tools recommended for the week were OpenAI embeddings, Chroma and a LangChain retriever; Chroma was chosen over options like AWS OpenSearch because it takes less time to set up.
• How you chunk the data decides whether retrieval works. Eremenko's tip is to use overlap, chunk by document section and chunk with context in mind, rather than cutting at fixed sizes such as 500 characters or tokens. His reasoning is that embedding models are shallow. His example: if "employee of the year was awarded to" lands in one chunk and the name in the next, the system can't find the answer.

2025-08-30 · open the original ↗

• "Chat" models are fast thinkers; reasoning models think slowly. The host disputes the label "chat model," since both kinds are used in chat. He maps the split onto the fast and slow thinking systems that Daniel Kahneman, Amos Tversky and others studied from roughly the 1960s and 70s. Models like OpenAI's 4o, 4.1 and 4.5 emit tokens as a stream of consciousness with no reflection. Reasoning models, starting with o1 about a year before recording, iterate over steps and check their assumptions, like someone working a problem out on a whiteboard.
• Reasoning trades compute for sharply lower error rates. The host says checking each step before answering can produce far more accurate, nuanced and complex solutions without a larger model. The main cost is computation. His evidence is that models from both OpenAI and Google recently earned gold at the International Math Olympiad using reasoning. Eremenko notes that the first week of the bootcamp has participants explore when to use each kind.

2025-08-29 · open the original ↗

• AI use is a sliding scale, not abstention. One speaker calls himself "a bit of an AI vegan", meaning he chooses what AI he will accept, as opposed to online critics who reject it as unethical "in any flavor". The other says he is in the same place: he uses AI tools regularly for programming while holding "a lot of concerns", and notes that people tend to hear "vegan" as total abstention.

2025-08-29 · open the original ↗

• A FreePBX zero-day lets attackers in and run up phone bills. The host says it began hitting systems around August 21, 2025; Sangoma announced it five days later and released an edge-module fix the same day. It affects FreePBX 16 and 17, and he advises checking version 15 and older anyway, since attackers can also pivot to other servers behind the firewall.
• Only servers exposing the admin GUI to the internet are vulnerable. Under Connectivity > Firewall > Services, Web Management (and Web Management Secure) should be open to the Local zone only; Internet or Other toggled on is the danger sign. Trusted remote WAN IPs go on the Networks tab instead, and Intrusion Detection should show as running to block brute-forcing.
• Sangoma's patch prevents infection but does not clean a compromised system. Over SSH the host checks the indicators from a Bleeping Computer article: /etc/freepbx.conf missing or modified since August 21; a .clean.sh script in /var/www/html; modular.php in the Apache access logs; calls to extension 9998 in the Asterisk full logs; and a user named ampuser in the ampusers MySQL table.
• A compromised server should be restored from a pre-attack backup. Restore from a backup or snapshot taken before August 21, then apply the patches from Sangoma's forum post and check phone bills for unusual charges, especially international calls. Crosstalk, which sells the help, says its own PBX customers are likely unaffected because it never exposes the admin GUI except via VPN or authorized IPs.
• Upgrading off end-of-life FreePBX 15 means building a new server. Version 15 gets no more security updates after September 1, 2025, the host says, and version 17 runs on a different base operating system, so upgrading means building a new server, installing FreePBX and restoring the configuration rather than upgrading in place; Crosstalk offers to do the migration.

2025-08-29 · open the original ↗

• CrewAI organizes specialized agents around roles, tasks and handoffs. The host describes CrewAI as an open-source Python framework where each agent (researcher, planner, writer, engineer) gets a clear goal and its own tools. Projects are broken into tasks, and the framework routes work between agents through shared memory and guardrails. Autonomy is adjustable: a loose crew suits open-ended research, a deterministic workflow suits auditability or client agreements, and a hybrid has a flow call a crew at key steps.
• His three use cases keep a human approving the output. In code review, separate agents handle static analysis, test proposals and risk explanation, but humans still approve and merge. In content creation, persistent research, writer and editor roles keep the voice and style guide consistent across runs. In supply chain or customer support, a monitoring agent, a planner agent and a vendor- or customer-facing agent respond to exceptions "in minutes instead of hours," with every decision logged.
• Durable roles fix what earlier agent setups lacked. The host argues that a single agent with tools blurs roles and needs constant prompt juggling, and that early multi-agent experiments proved the idea but were unstable. CrewAI's explicit task breakdown and structured handoffs mean fewer loops, less brittle prompting and easier scaling to production. He adds that swapping models or tools is a configuration change, not a rewrite. He states that CrewAI does not sponsor him or the show.
• Multi-agent systems need review gates and limits on agent permissions. The host's list of precautions: define review gates, track sources, restrict tool permissions, monitor spend, and keep a human in the loop for consequential decisions. His advice for getting started is to codify roles you already play and split one weekly task into two or three agent handoffs.

2025-08-28 · open the original ↗

• Destructive APIs make API design matter more for LLMs. One speaker argues that an LLM "fumbling around in the dark" with local tools may be harmless, but plugged into APIs through Postman it could charge a credit card or make a mistake. The other agrees, adding that reaching for the wrong tool makes a job much harder and "the right tool for the right job makes all the difference."

2025-08-28 · open the original ↗

• Leading with value beats self-focus, even in insecure times. Depenbusch, who works in people analytics at Snowflake, argues that going into meetings asking "what's actually needed right now" rather than pushing your own recommendation makes you stand out, because almost everyone else goes in "me-centric". Hanselman countered that fear of layoffs makes self-protection natural, a leap "without a parachute"; she conceded it goes against human nature but said it is more likely to get you the job or promotion.
• Give value before asking for mentorship. She says she usually ignores bare "will you be my mentor" DMs, but a stranger who kept sending her tools, articles and thoughtful comments built the relationship he wanted; she reciprocated with feedback on his presentation and a free copy of her data-viz course. Hanselman says "can I pick your brain" has become actively irritating to him because it is unfocused and disregards the other person's time.
• Be a thought partner, not merely a task completer. Depenbusch says value means asking what the goal is (her example: proposing a better attrition analysis than the one requested) while sometimes doing the hour-long data pull anyway to build rapport and earn the credibility to push back. When overruled, she says, get on board: "start climbing that mountain" and find out sooner whether it is the wrong one.
• Against idea theft, rely on relationships and documentation. Answering Hanselman's question about less friendly workplaces, she says win allies one by one who credit you in rooms you are not in, and put proposals in writing so you can show your boss an idea implemented 6 months later. Her own case: weeks of unrelated analysis for another Google team in her first year led to an offer to join it 6 months later; Hanselman calls this the compound interest of relationships.
• Value can be informational, emotional or strategic. Information (tools, resources) and emotional value (encouragement, feedback, advocacy, humour) are open to anyone; strategic value (perspective, surfacing risks, better questions) grows with expertise. She also reports from her dissertation that work passion mirrors romantic passion, with a honeymoon of about 6 months, and that once she became a parent her priority became a job that lets her leave at 4:00.

2025-08-28 · open the original ↗

• Free UniFi filtering gives a basic adult and malware filter only. Dave Barger, presenting, shows that any modern UniFi gateway includes ad blocking, forced safe search and manual allow/block domain lists for free, but the filter itself is limited to off or "basic"; per-category filtering is grayed out until you subscribe.
• Paid CyberSecure unlocks category blocking with maintained definitions. CyberSecure Enhanced, by Proofpoint and Cloudflare, costs about $99 a year and adds the full category list plus IDS/IPS options; blocking photography and shopping categories made B&H unreachable in his test. Part of what you pay for, he says, is keeping those categories up to date.
• Built-in filtering beats third-party DNS filtering, Barger argues. Pi-hole's free blocklists mainly target ads and trackers, and commercial DNS filters are typically priced per user and can't be managed from the UniFi console; UniFi's version keeps everything in one place and relies on more reputable companies for definitions.
• The policy engine blocks specific apps free, with schedules. After a UniFi update that arrived before release, app blocking moved under the Policy Table or the new object-based networking; he builds a rule blocking TikTok on the default network, with an optional schedule. He sees UniFi slowly closing the gap with full UTM devices such as a FortiGate.

2025-08-28 · open the original ↗

• Many problems pitched as "agent" problems need no LLM. Eremenko calls the bootcamp's first week the "mindset shift" week. His view is that executives swayed by the hype want an agent, or generative AI, where a simple LLM call, a plain script or robotic process automation (he names UiPath) would do. The first step is to reframe the problem with business stakeholders.
• Judge models against your own business benchmark. Participants compared 13 LLMs and contrasted reasoning models with chat models. Eremenko endorses a point he credits to a recent guest on the show: public benchmarks help, but a business needs its own benchmark.