tech

2025-03-28 · open the original ↗

• Baseball analytics has gone from Moneyball experiment to AI arms race. The host says every MLB team now employs data scientists, and that Statcast, introduced in 2015 with high-speed cameras and radar in every ballpark, generates seven terabytes of data per game, far more than human analysts can sift.
• Machine-learning scouting finds players that traditional stats undervalue. Models ingest college and minor-league stats, biomechanics and video, and weight metrics such as exit velocity, launch angle and spin rate; the host's example is a prospect with a low batting average who consistently hits the ball hard. Teams also use the analysis to tailor drills, such as for a hitter who struggles with high fastballs.
• In-game decisions now lean on simulated matchups and positioning models. Managers consult win-probability tables and pitcher-versus-batter predictions on tablets in the dugout, and data-driven defensive shifts proved effective enough that MLB introduced rules limiting extreme shifts. The host stresses the tools augment managers rather than replace them.
• AI is reaching fans and officiating as well. Broadcasts show live win-probability graphs and catch probability on replays, and the host reports that during 2025 spring training MLB let players challenge human umpires' ball-strike calls to an automated system, which drew mostly positive reviews.
• Any traditional industry can be transformed by pairing domain expertise with AI. The host's lesson is that if America's oldest major sport can reinvent itself, finance, healthcare, manufacturing and retail can too; early adopters gain an edge until others catch up. He forecasts AI won't replace product managers, physicians or supply-chain planners for the foreseeable future, but will arm them with better information.

2025-03-27 · open the original ↗

• Decarbonisation has progressed further than its advocates' gloomy messaging admits. Campbell cites an argument that the world is about a third of the way to non-carbon-emitting energy; the Kyoto-era projection of a four-degree rise is "off the table", though he says the 1.5-degree target will be missed and the outcome is somewhere between two and two and a half degrees. Straight-line projections look worse than real curves, which twist and accelerate.
• Wind and solar are now cheapest, but not baseload power. Measured by levelized cost of energy, he puts wind at about $25–26 per megawatt-hour and solar at $23–24, against roughly $40 for combined-cycle natural gas. Because they generate only when wind or sun allow, rooftop solar in New South Wales now destabilises the grid at midday, and Sydney's power company can disconnect a home's panels; he says the real change of the past 20–30 years is grid command and control moving from top-down to two-way.
• Adoption is hitting a wall without storage, best done by pumped hydro. He says grids cope with 10–20% rooftop adoption but going higher needs more storage and transfer; residential solar is expensive per unit of power, with a 5–20 year payback depending on electricity prices. At grid scale he names pumped hydro as the cost-effective store, citing an Australian university study that found many hills with water at the bottom where a reservoir could be dug at the top.
• Goals are well behind schedule, and costlier money made baseload harder. Data centres draw tens of megawatts around the clock and need baseload; he points to Microsoft's Quincy sites beside two hydroelectric plants. Gas, nuclear and coal plants must be fully built before earning anything, which recent inflation and interest rates punished, whereas a solar farm can connect its first 20 of 1,500 acres, or a wind farm its first 10 or so turbines, and grow from revenue. He warns corporate claims involve measurement games and "greenwashing".
• Nuclear stalled over weapons proliferation more than over accidents. Campbell says light water reactors won because they bred plutonium — about 200 pounds from a 50-ton assembly over two years — and that the US has not reprocessed fuel since the Ford administration stopped it, while France does. Newer passive-shutdown designs cool themselves without pumps; by his account Fukushima units 1–3 failed because they lacked it and their diesel generators were in flooded basements. The two Vogtle reactors in Georgia, operational in 2024, were the first new US reactors in 40 years, but with no further plans he expects the skills to decay; he suggests the US might need about 20 built over 10–15 years.

2025-03-27 · open the original ↗

• Step-by-step reasoning models could scale to many Nobel-level minds. The host, prompted by OpenAI's o1 release, argues synthetic data could broaden reasoning beyond math, chemistry and code, and forecasts that within "a matter of years" a data center could hold a million Nobel-prize-level machines, the "powerful AI" of Anthropic CEO Dario Amodei's essay.
• Relying on LLMs for cognitive work makes people dumber. Monbiot cites studies showing users' cognitive capabilities decline when LLMs do their thinking, and warns of a "great intelligence divide" in which a few build ever more powerful AI while everyday users degrade. Her focus is on how AI can augment human agency rather than on alignment.
• AI's real payoff is freed capacity for ingenuity, not a modest speed-up. She calls the roughly 15% average performance gain for white-collar workers found by Ethan Mollick and others uncompelling, and prefers Andrew Ng's framing of the 10x engineer as exponentially more effective because AI frees time and mental capacity.
• Used well, AI expects more of people, not less. Monbiot cites the former CEO of Drift building an AI-agent-first customer-success startup whose first hire after engineers was a lawyer who also oversees HR. She argues people must think laterally and drop ego: engineers whose work is democratized, and non-technical people who can no longer claim "I'm not a coder".
• The "anyone can now code" narrative isn't fully true yet. Monbiot, a linguist by training, says she got halfway through a Python course, tried building, and had to go back and learn more. The host argues the best learning pairs a specific goal with just enough fundamentals, with the balance differing by person.

2025-03-26 · open the original ↗

• Small changes to a technology like language can reshape civilization. Monbiot cites a scholar's argument that the ancient Greeks' addition of vowels made expression more fluid and helped language spread, contributing to the birth of democracy, an impact only visible in hindsight. She argues AI's effect on humanity is similarly impossible to predict yet but probably huge.
• Advanced language arrived long after brains capable of it, the host says. He puts sophisticated language at only tens of thousands of years old while saying Homo sapiens have been genetically "pretty much the same for millions of years", and tells listeners to double-check that fact rather than quote him.

2025-03-25 · open the original ↗

• Humans should be AI's meaning makers, not lemmings off a cliff. Monbiot traces her recent writing to the book God, Human, Animal, Machine, which she recommended on her previous appearance (episode 823), and to an op-ed in the Polish magazine Przekrój asking whether we are ceding humanity to machines. Her answer: humans created the opportunity and the mess, and are good at cooperating to fix messes.
• AI is part of human cognitive evolution, as language was. Drawing on John Vervaeke's lecture series The Meaning Crisis, she calls language a "psychotechnology" that enabled self-reflection and cites his argument that Greek vowels helped give birth to democracy. AI is part of us the way language is, but not all of us, because humans are embodied and embedded in the world.
• Leaning on LLMs risks a "great intelligence divide". Monbiot cites studies that users' cognitive capabilities decline when LLMs do their thinking, and dismisses the roughly 15% average gain for white-collar workers as uncompelling. She favours Andrew Ng's 10x-engineer framing, and the example of an AI-native startup whose first post-engineer hire was a lawyer also running HR, as signs AI should demand more of people.
• LLM drafting drags writing toward the mean. Monbiot says the more iterations she does with an LLM, the more her "aha" insights get averaged out and she has to reinsert them, so she is unsure it saves her time. She and the host agree the states worth optimizing for are flow and peace; the host argues free, ad-funded platforms are built to keep users in a frenetic, fearful state instead.
• Everyone will have AI twins working on their behalf, she predicts. Monbiot's "virtual human economy" thesis drives her advisory work for a startup offering vetted experts AI twins that can earn consulting income in any language, and a custom twin of a healthcare CEO deployed in Slack to carry company culture. The host describes his own early idea for an embodied AI that would steer him back toward flow and peace.

2025-03-24 · open the original ↗

• Build AI strategy on pillars and objective truths, then derive tactics. Alake, who leads and mentors AI teams at MongoDB, calls this the year of "show me the money" for AI and says his method is first-principles: identify the pillars around a persona, then the things within each pillar that will not change short of a platform shift. His example for a developer-focused strategy is GitHub as a pillar, asking what developers actually do there, then forming convictions, assumptions and tactics from that.

2025-03-23 · open the original ↗

• This is the "multi" era, not just the year of agents. Alake calls 2025 the "multi" era: agents went mainstream this year, so leaders should look further ahead: to multi-agent architectures, multimodal embeddings covering text, images, video and audio in one space, multilingual embeddings, multi-step reasoning, and multiple retrieval mechanisms, since "vector search is not all you need".
• Voice agents will be widely built but won't dominate, he predicts. Alake expects everyone to experiment with voice interfaces given AI's experimental nature, while holding a strong opinion that voice will not become the main way people engage with agents.
• Memory management will become a central agent-design problem. He recounts that Andrew Ng's first question to him, in April of last year, was how to model memory in agentic systems, and argues agents will need several forms of memory and recall, as human brains do.

2025-03-22 · open the original ↗

• The AI stack is described differently to executives and to developers. Alake, whose blog post on AI stacks the host says was the top Google result for the term at recording, gives executives three layers: application (e.g. Cursor), tooling (e.g. MongoDB) and compute (e.g. Nvidia). For developers he breaks it into more layers, including programming language, model provider and database.
• Choice of programming language constrains the rest of the AI stack. Alake stresses that many libraries further up the stack exist only in Python or only in JavaScript, though some are evolving to support both, so the language decision comes early and matters.

2025-03-21 · open the original ↗

• ClearlyIP has opened a community forum for an unexplained TangoPBX. The forum is live at community.tangopbx.org; the host, who calls himself a big fan of ClearlyIP, declines to say what TangoPBX is and calls the forum the first in a series of announcements due over the coming months. He is "Crosstalk" on the forums.
• The host says FreePBX's own forums have driven users away. He attributes this to a noticeable lack of interest and engagement from Sangoma, FreePBX's parent company, and to heavy-handed moderation, which he says caused an exodus. He presents the new forum as ClearlyIP's attempt to rebuild a welcoming open-source PBX community, which he calls desperately needed.

2025-03-21 · open the original ↗

• This upload holds a different episode than its title names. The video posted as 872 is a copy of episode 870 on OpenAI's Deep Research: both run 17:10, their captions are identical, and neither mentions the Majorana 1 chip. Episode 870's summary, dated 2025-03-14, covers what the video says.

2025-03-20 · open the original ↗

• MongoDB Atlas exists because no developer wants to be a database administrator. Alake describes Atlas as a managed, distributed MongoDB running across AWS, Azure and Google Cloud in multiple regions, bundled with streaming and dedicated search nodes.
• Vector spaces place similar meanings close together. The host explains embeddings from Tomas Mikolov's word2vec, around 2011–2012, which learns a word's meaning from surrounding words without labels; in a trained space, spelling variants coincide, NoSQL sits near MongoDB, and dimensions can run to hundreds or thousands. Documents embedded the same way let RAG systems retrieve relevant items from, say, a million documents by a query's meaning.
• MongoDB folded search and vector search into one database. Alake says customers had been bolting search engines, extensions or separate databases onto MongoDB, so it added Atlas Search and then vector search, allowing lexical, vector and hybrid search in a single database where he has seen some teams run two or three.

2025-03-19 · open the original ↗

• You can message anyone if the question is precise and theirs to answer. Hanselman recounts messaging a distinguished engineer he had never met about how the ONNX runtime does quantization for an NPU, which led to a half-hour conversation. The senior PM watching said they had never felt empowered to do that. Hanselman concedes his VP title and name probably played a part, not only the quality of the question.
• Reputation is built by never bothering people with randomness. Mark says his title, seniority and reputation let him contact anyone. He leaves requests that could have been Googled, such as "where's the deck", read and unanswered, but answers thoughtful ones; someone at the last Ignite told him he had replied to their email 15 years earlier. Hanselman compares this to Donald Knuth's reward checks for finding errors in his books: people do their homework before approaching him.
• Answer a good question once, publicly, and send the link. Hanselman says he turns a thoughtful question, such as an Indiana University student's email about impostor syndrome, into a blog post or other public piece and replies with the URL, so others benefit too.
• Admitting confusion is fine only after you've done the homework. Mark says he forces himself to say "I don't understand" in meetings, but doesn't want to stop a meeting over something he should have prepared. Hanselman draws the line between not having prepped and having done the research while some points have not yet clicked, which he says is appropriate at every level.
• At senior levels, the right point can land in the wrong venue. Mark weighs whether an opinion is deep enough to raise in a meeting with the senior leadership team and Satya. Hanselman says his VP replied "not helpful" when he made a point to 83 people that only three needed. He keeps his emails to Satya to two or three a year, and warns that asking basic questions flips the "bozo bit", which is hard to flip back.

2025-03-19 · open the original ↗

• MongoDB's document model mirrors how developers think. Alake contrasts relational databases' tabular rows and columns with MongoDB's key-value, JSON-like documents, and says seeing data represented the same way in the application and data layers made him a full-stack MEAN developer (MongoDB, Express, Angular, Node.js) after he had disliked programming.
• Flexible schemas suit fast-moving AI experimentation. The host says his startups chose MongoDB around 2014 because pages' data needs were uncertain; Alake argues AI work is the same, with schemas evolving during experimentation and lockable once stable in production. The host's example is storing extra output fields from a Llama model that the OpenAI API lacks, without a schema migration.
• JSON unifies the AI stack from LLM to database. Alake argues that with LLMs now offering JSON mode and structured output, JSON runs through the application layer, the data layer and the model, and that avoiding schema migrations matters because developer speed is a competitive advantage.

2025-03-19 · open the original ↗

• VTM is a text-mode desktop that deserves far more attention. Hanselman, who says he doesn't know its developer, notes the open-source project had about 2,000 GitHub stars and argues it deserves 20,000. He likens it to DESQview, the old text-mode multitasking environment, and places it among TUIs such as Midnight Commander and Malcolm Smith's YEdit. He says the whole thing is about five megabytes.
• Its windows behave like a desktop, with tmux-style splits. In the graphical mode he shows resizable windows with z-order that can be flung around with physics, running DOS, bash, PowerShell and Vim side by side. There is a split window manager with "empty slots" to drop terminals into, hotkeys such as Shift+Alt+W, and an XML file configuring the five launcher apps. He notes it "gets a little angry" with his three-monitor layout.
• Run as a text app, it reattaches to detached sessions on an infinite canvas. After closing the graphical version, Hanselman starts VTM inside Windows Terminal and reconnects to the still-running processes, as screen and tmux do. Viewport markers show how far away off-screen windows are, which he likens to an infinite Canva-style terminal manager. He calls it "pretty obscure and confusing", with distinct left, right, middle and Ctrl-click behaviours.
• Windows has long been "missing the text mode boat". Hanselman points back to his blog post of that name from 2004, 21 years earlier, and to Gui.cs for building console apps with UI. He credits both the Windows Terminal team and VTM's developer for this working inside Windows Terminal.

2025-03-18 · open the original ↗

• RAG grounds an LLM in facts by two retrieval patterns. Polak, who heads developer advocacy at Confluent, distinguishes term retrieval, long available in Elasticsearch and OpenSearch but requiring the user to know the exact term, from embedding retrieval, which searches vectors by similarity so the user need not.
• An agent is given a goal and tools, not steps. On start-up it discovers its tools (query templates, databases, web search, logs), a planner LLM decides what to run, and a judge agent checks the result. In her example of projecting three months of sales for a fashion collection, the judge sends the answer back because it ignored marketing-campaign data. Triggers need not be human: a build error can start a code assistant's agentic process.
• Agents are microservices and need an event broker. She argues production agent systems face the old microservices problem of lost requests, solved by an immutable log such as Apache Kafka. Apache Flink then filters, joins and enriches events in real time — for instance dropping build-error events for a user who has switched that help off — and its stateful processing plus Kafka's storage can serve as the agents' short- and long-term context.
• Exactly-once processing matters more because LLMs are non-deterministic. Temperature makes the same question yield different answers, so a duplicated event can produce repeated or contradictory suggestions. Flink's external two-phase commit protocol gives exactly-once guarantees across Flink and Kafka.
• Confluent is adding agent-oriented features, some proprietary. Polak says the two-phase commit on both Flink and Kafka "works really, really well", and Flink now exposes richer exception context. Confluent's proprietary Flink AI lets customers register a model (such as OpenAI models on Azure) and use it as a transformer inside a Flink job, for repetitive tasks like transcripts or SQL generation, without a separate microservice.

2025-03-18 · open the original ↗

• A flexible document schema suits AI's constant experimentation. Alake, a staff developer advocate for AI/ML at MongoDB, argues the JSON-like document model lets the data layer evolve while requirements are unclear and be locked once in production, avoiding schema migrations; with LLMs offering JSON mode and structured output, JSON now unifies application, data and model. The host, who says MongoDB does not sponsor the episode, gives the example of storing an extra output field from a Llama model that the OpenAI API lacks.
• One database can handle operational, lexical, vector and hybrid search. Alake says customers had bolted search engines, extensions or separate vector databases onto MongoDB, sometimes running two or three databases, so MongoDB built Atlas Search and then vector search into the managed, multi-cloud Atlas service. The host adds that a single vendor avoids version mismatches between linked systems.
• Store an agent's tools in the database; retrieve only a few. Alake's open-source library Memorizz implements a "MongoDB as a toolbox" pattern: since OpenAI's guidance is to put only 10 to 20 tool schemas in context, store perhaps a thousand tools with embeddings of their docstrings, then fetch two or three before each LLM call. He jokes that stuffing a million-token context window on every call depends on "how deep your pocket goes".
• The "multi" era is here, and memory is the next frontier. Calling 2025 the "multi" era, Alake expects, beyond single agents, multi-agent architectures, multimodal and multilingual embeddings, multi-step reasoning and multiple retrieval mechanisms, since "vector search is not all you need". He predicts voice agents will be widely tried but not dominant, and that memory management, as pursued by MemGPT's maker Letta and the HippoRAG paper, will become central.
• Ground AI strategy in objective truths; test tactics in the market. Alake's "arena framework" pictures a coliseum whose pillars are segments of objective truth (for developers, GitHub and what they always do there), with tactics battling it out in the market. For teams he prescribes being aggressive, centralized, coordinated, resourceful, efficient and data-driven.

2025-03-17 · open the original ↗

• Extrapolate current curves a little rather than predict AGI. Godbole, recently retired from Google, says his post "From Knowledge to Wisdom" starts from observable trends: each year models get much better while inference cost for the same functionality falls by orders of magnitude, context windows grow and latency improves, as with Tesla's Roadster preceding cheaper models. He notes there is no consensus definition of AGI.
• AI coding tools will produce "full-stack employees". Because tools such as Cursor, Lovable, Replit and Copilot make software accessible to non-engineers, Godbole expects a PM to arrive with generated mocks, a prototype and early user research before an engineer spends a second on it. He suggests roles are bundles of skills that may be unbundled and dynamically rebundled, leaving the shape of future organizations unclear.
• Cheap reproduction could make software resemble the music industry. If an undifferentiated SaaS product can be cheaply copied, Godbole reasons through a chain of "ifs" that most offerings face near-perfect competition while a few highly differentiated winners take outsized returns. He expects the effects to be discontinuous, such as small teams simply not hiring a PM.
• When knowledge is commoditized, wisdom becomes the differentiator. Godbole argues good ideas still come from humans, and that cultivating wisdom resembles cultivating curiosity, open-mindedness and adaptability. His aim is not more agentic AI but using AI to make himself more agentic, self-authored and pro-social, which he says also improved how he wields models, since language makes human assumptions the models do not share.

2025-03-16 · open the original ↗

• ASP.NET Core offers three authorization models; policies scale best. Balliauw separates authentication (proving who you are, e.g. via the built-in cookie handler that sets an encrypted cookie) from authorization (what that identity may do). Role-based checks are simplest; claims-based checks add properties but bloat the cookie in complex systems; policy-based authorization lets you write custom requirements, such as a database lookup, and combine policies.
• OpenID Connect extends OAuth for proving login to APIs. In a microservices setup the client sends the user to an authorization server, which returns a signed token the client passes to each API; he says OpenID Connect adds the token format, signing mechanisms and refresh rules on top of OAuth. Because ASP.NET Core ships the client side, it works with any compliant server, such as Duende IdentityServer or Entra ID. He recommends Microsoft's ASP.NET Core authentication docs to start, and reading the OAuth and OpenID specifications to understand the design choices.
• IdentityServer is an SDK for building your own identity provider. Balliauw, head of customer success at Duende Software, describes its product as a way to "build your own Entra ID". His reasons to use it: reusing an existing user store, avoiding a hosted SaaS service, plugging custom logic into every step of the flow, and configuring industry profiles such as the financial-grade API's proof of possession.
• Single-page apps should not hold tokens in the browser. He says a token in local storage can be fished out by anyone with some IT background and used against the API, and proof of possession does not help because its key would also sit in the browser. The backend-for-frontend pattern, which Duende also sells as BFF, keeps tokens on the hosting server, which proxies API calls and handles refresh, while the browser uses only a cookie — more secure and, he says, a much simpler client programming model.
• Dynamic grants and passkeys are the changes he is watching. A grant-management spec in progress would let users add or remove consented access, such as a second bank account, without discarding existing tokens. He is "a big fan" of passkeys: a per-site key pair on a trusted device means no stored passwords, nothing useful in a breach "at least until quantum computers become a big thing", and revocation by deleting a public key.

2025-03-16 · open the original ↗

• LLMs have turned much of machine learning into prompting plus evals. Godbole, a core member of Google's Gemini team until he recently left, calls LLMs an "arbitrage of ML talent": where adding intelligence to a product once meant PyTorch or JAX and data collection, it now means articulating the behavior you want in clear English and measuring it with good evals, a very different skill set.
• The mental model in his prompting playbook outlasts its prescriptions. Godbole says the second half of the playbook he published before leaving Google, examples of good and bad prompts, reflects today's models, since prompting o1 or Gemini Flash Thinking feels different from GPT-4o or Gemini 1.5. The first half, on pre-training versus post-training, he considers more future-proof.
• Pre-training learns every "cinematic universe"; post-training picks one. He argues that for an LLM there is no objective fact: "Bilbo Baggins is a hobbit" is true only in Tolkien's universe, and web text approximates the union of all such universes. Post-training narrows the model to an assistant-and-user universe, so post-trained models are statistical models role-playing as human raters role-playing as AI assistants.
• Prompt as if briefing an all-knowing stranger who doesn't know you. Godbole says complaints about why a prompt is failing usually work better when pasted into the prompt, and that models respond best to extremely explicit, low-context communication. He recommends the book Nonviolent Communication, learned in couples therapy, because being explicit about the observable behavior you want works on people and models alike.

2025-03-15 · open the original ↗

• Google bet on ML code completion before ChatGPT made it obvious. The host points to a July 2022 Google Research post, "ML-Enhanced Code Completion Improves Developer Productivity", noting that at the time such tools might have seemed more trouble than their errors were worth. Godbole, who moved from medical imaging to this work, says early results were promising and that steadily improving hardware and model size made engineering productivity a natural use case for sequence-to-sequence models.

2025-03-14 · open the original ↗

• Accessibility matters to everyone because disability is often temporary or situational. Tran, a Microsoft partner technology strategist and 12-year stroke survivor with low and double vision, cites about 1.6 billion disabled people, roughly one in four adults, and distinguishes permanent (blindness), temporary (an eye infection) and situational (sun glare) disability. He says Microsoft built its tools in at the design phase, and that he first assumed he could not do his job until he found them.
• Built-in speech tools work both ways and help anyone multitask. Windows Dictation (Windows+H) turns speech into text, which he uses because left-side weakness makes him a slow typist; Narrator reads text aloud for his low vision. Read Aloud, in Microsoft 365 apps and Edge, reads email or web pages and, he says, in the sender's accent — an Australian trainer's email came in an Australian voice.
• Small display settings make the screen trackable. He enlarges the mouse pointer (to size 4; Giard uses 7 to find it across multiple monitors), the text size, or overall display scaling (he runs 175% on a 27-inch screen), and mentions high contrast, a blinking cursor and Magnifier for zooming into part of the screen.
• Microsoft also makes adaptable hardware. He points to the Microsoft Accessible Mouse, which takes add-on parts to fit the user's hand, and the Xbox Adaptive Controller, which lets people who could not use a standard controller play.
• Dictating to Copilot lets you prompt AI like talking to a person. He dictated a long, unproofread prompt for a three-day Seattle and Microsoft-campus itinerary; he says Copilot, which he says runs GPT-4o, interpreted it despite his atypical speech, then revised the plan when he replied conversationally. To find accessibility tools, he suggests asking Copilot or searching "Windows accessibility".

2025-03-14 · open the original ↗

• OpenAI's Deep Research does hours or days of literature review in minutes. The host, who notes OpenAI does not sponsor him, explains that it breaks a query into sub-tasks, searches the web for each, and iteratively synthesizes a cited report, pivoting as it learns. OpenAI trained it with end-to-end reinforcement learning on browsing and reasoning tasks, and it can read uploaded files, plot with Python, embed images and cite specific passages.
• Deep Research jumped far ahead on Humanity's Last Exam. On the 3,000-question benchmark spanning over 100 subjects, the host's table at recording shows GPT-4o at 3%, Grok 2 and Claude 3.5 Sonnet around 4%, Gemini Thinking about 6%, o1 and DeepSeek R1 around 9%, o3-mini (high) about 13%, and Deep Research at 27%. He predicts that, as with past coding and math benchmarks, the exam won't take long to be conquered once AI gains traction.
• In the host's use, it asked sharp questions and saved hours. Drafting the syllabus, title and abstract for his four-hour agentic AI workshop at ODSC East, he supplied partial modules and past abstracts; it asked about audience level, hands-on coding and tone, then spent three minutes across eight sources, including his own website, to produce a draft he only lightly edited.
• Google's and Perplexity's versions trade depth for other strengths, the host judges. He says Google's Gemini-based tool shows a user-guided research plan first and is reliable but less nuanced, while Perplexity's is fast and free but shallower; for mission-critical analysis he gives OpenAI's the edge despite its cost.
• The catches are cost and occasional overconfidence. The host pays $200 a month for Pro, which includes 100 Deep Research queries a month, and warns it can hallucinate or present rumor as fact, though he has caught neither and says OpenAI reports markedly lower hallucination rates. He forecasts efficiencies such as smaller models like o3-mini will bring more queries to all paid users and eventually free access.

2025-03-13 · open the original ↗

• Regulatory economics, not science, keep novel devices out of medicine. Jepsen cites a paper averaging $658 million and 13 years to approve something novel and therapeutic, built on "quality builds" of about 10 units that rule out Moore's law and contract manufacturing. She says a million papers in 20 years on infrared, ultrasound or electromagnetic treatments have produced a rounding error to zero in the healthcare system. A single rare disease such as glioblastoma, about 5,000 US patients a year, ends up as a treatment costing around $1 million after roughly $1 billion spent.
• Openwater sells an open-source, general-purpose device as its business model. Jepsen likens it to Android, where the "apps" are regulatory approvals her customers obtain. She says the system costs $10,000 now, against $1 million earlier, and could reach the cost of a smartphone at volume. It is released under AGPL and Creative Commons 4.0 ShareAlike; the company has raised $54 million last year and $100 million in total, and expects to ship a few thousand units this year and go profitable. A $50 million non-dilutive check from Ethereum's founder came with a veto over ever closing the source.
• Smartphone camera chips and a pulsed laser can spot strokes. Jepsen says recording holograms with the micron-sized pixels of phone cameras gives blood-flow readings in phantom tissue accurate to 0.1%, which she calls 20 times better than CT, MRI or Doppler. She describes it as an "ECG for your forehead" to route large-vessel-occlusion stroke patients, a type she says kills 6.5 million a year, to the 5% of hospitals that can treat it. Treated within two hours, she says, there is a 90% chance of no neural deficit; the laser starts production next quarter.
• Low-intensity ultrasound tuned to cancer's resonant frequency kills it selectively. Jepsen reports sweeping frequencies on 16 glioblastoma lines grown as organoids, then moving 30 mice into remission with no damaged healthy cells found at autopsy. She says the treatment releases proteins that "vaccinate" the body against its own cancer. In a severe-depression study almost half the patients went into remission on 10-minute doses over three weeks, and 150 kHz pulses at a 10% duty cycle break microclots below five microns in preclinical work. None of it is FDA-approved yet.
• Open safety data and hacker access are the next bottleneck. Jepsen wants all safety and outcome data shared with regulators, doctors and patients. She notes that ultrasound use on people is regulated as Class 2, that experiments on anything with a backbone need ethics review boards, and that the open-source community needs hardware and lab space. Hanselman, whose team runs Microsoft's open source programs office, offers Microsoft Reactor spaces.

2025-03-13 · open the original ↗

• The Deep Learning Tuning Playbook aims to turn alchemy into chemistry. Godbole, its first author, says it pooled his Google Brain co-authors' decades of experience into a systematic, architecture-agnostic process: set up a good initial experiment, then decide each next experiment from the data collected, given a compute budget and constraints. He credits the co-authors as the real brains and calls it their "two cents", not the final answer.
• Releasing it as a forkable markdown file made it a community resource. Rather than a PDF on arXiv, the team published it on GitHub under a permissive license to invite forks and pull requests; community members crowdsourced translations the authors do not endorse. The host reports 28,000 GitHub stars at recording and says it answers the question his deep learning students always ask about why to add layers or neurons.

2025-03-12 · open the original ↗

• A deep learning system beat radiologists at predicting breast cancer from mammograms. Godbole, co-first author of the 2020 Nature paper "International evaluation of an AI system for breast cancer screening", says it predicted biopsy-confirmed cancers two or three years ahead, depending on the country's screening guidelines, better than expert radiologists. It was benchmarked on a large retrospective US and UK dataset of what he recalls as a hundred-and-some thousand patients, and beat every radiologist in a separate reader study.
• The bigger promise is combining AI and humans and widening access. Godbole says humans and AI have different strengths, so their synthesis can be powerful, though he is out of date on that literature. He highlights countries with many patients per radiologist, where AI could substantially improve access to potentially life-saving screening.

2025-03-11 · open the original ↗

• The new U7 Pro XG and U7 Pro XGS are fanless redesigns. Both have a near-flat face and an all-metal back built to shed heat, so no fan and no noise, plus a 10 Gbps RJ45 port powered over PoE+. The new mounting plate is not a drop-in replacement, but it keeps the old screw-hole pattern, so no new holes are needed.
• The XGS is the higher-capacity model, with a dedicated scanning radio. The host gives the XG as 2x2 MU-MIMO on all three bands, about 1,500 sq ft and around 300 clients. The XGS is 4x4 MU-MIMO on 5 GHz (2x2 on 2.4 and 6 GHz), about 1,750 sq ft and over 500 clients, and scans continuously for interference to adjust channels in real time.
• The host calls them a smarter buy for futureproofing. He describes the XG as essentially a U7 Pro and the XGS as a U7 Pro Max, each with a 10-gig NIC and the new fanless form. He places them between the existing U7 and the higher-end E7 lines, and says the 10-gig NIC alone justifies the small extra cost. Prices were shown on screen only, not spoken.

2025-03-11 · open the original ↗

• Copilot just shortens the search, Stack Overflow, copy-paste loop. Conery, who works on developer outreach for Microsoft's VS Code team, answers the "garbage code" complaint by noting that Stack Overflow answers need the same human vetting; trouble starts when Copilot is asked to do too much. Giard adds his own view that imperfect code getting you 80% of the way is still a big saving.
• Custom instructions pre-prompt every request and shrink your prompts. A markdown file of instructions in a .github folder is prepended to every request in edits, chat and inline chat. Conery keeps it a terse bulleted list of standards (two spaces, brace placement, class naming) and adds example code in fenced blocks, such as a Sequelize user class or nested test blocks, which Copilot follows; output stays non-deterministic. Detail moved into the instructions means a prompt can be as short as "make the data access layer".
• Edits plus strong instructions get about 95% there, in small steps. Edits changes files across a working set and shows per-change diffs to accept or reject. In his demo it turned a five-item spec into a SQL schema, which he then used itself as an instruction to generate a Sequelize data access layer "nearly perfect", missing only a file reference or two. He credits his boss Burke Holland with the rule of very small steps, and admits he could retry his demo until it worked.
• Agent mode runs steps in series and corrects its own errors. Then available only in VS Code Insiders, which updates as often as twice a day, it took his prompt to initialise a Node project, install packages and create a GitHub repo with the GitHub CLI, pausing for "Continue" at each step. He says it installs a missing CLI itself and fixes compiler errors in typed languages such as C# or TypeScript; he had no release timeline.
• Next edit suggestions propagate a change through the rest of the code. Renaming a Point class to Point3D prompts a diff adding a Z coordinate and updating later calculations; Holland's video shows it carrying a string changed to an enum through C# code. Conery was unsure whether it had reached regular VS Code yet.

2025-03-11 · open the original ↗

• Use AI to make yourself more agentic, not to build agents. Godbole, recently retired from Google, argues that once knowledge is commoditized the differentiator is how wise, pro-social and self-authored a person is. He wants AI to support metacognition and a reciprocal loop: acting more, hitting one's limiting beliefs, and growing. He credits Toronto professor John Vervaeke's work on wisdom and relevance as a major influence.
• AI predicted breast cancer from mammograms better than radiologists. Godbole was co-first author of the 2020 Nature paper showing a deep learning system predicted biopsy-confirmed cancers two to three years ahead, depending on the country, on a US and UK dataset of what he recalls as a hundred-and-some thousand patients, and beat every radiologist in a separate reader study. He sees the bigger promise in combining human and AI strengths and in expanding access where patients far outnumber radiologists.
• An LLM must be told which "cinematic universe" it is in. Godbole argues that for an LLM there is no objective fact: pre-training fits the union of all universes in web text, and post-training narrows it to an assistant-and-user universe, so models are statistical models role-playing as human raters role-playing as AI assistants. This first half of his prompting playbook, he says, will outlast its prompt examples.
• Prompt as explicitly as you would brief a stranger. He advises imagining a person off the street with all human knowledge but no knowledge of you, and pasting one's complaints about a failing prompt into the prompt itself. Skills from couples therapy and the book Nonviolent Communication carried over, and he reports an anecdotal inverse correlation between ML expertise and prompting skill.
• Cheaper, better models will unbundle job roles and squeeze undifferentiated products. Extrapolating modestly from current curves, Godbole expects "full-stack employees" such as PMs who arrive with prototypes and user research before engineers are involved. Because copying an undifferentiated product will be cheap, he speculates software could come to resemble the music industry's winner-take-most economics, with discontinuous effects such as small teams not hiring PMs.

2025-03-11 · open the original ↗

• Distilled DeepSeek R1 models run locally on a Copilot+ PC's NPU. Hanselman uses the free AI Toolkit extension for Visual Studio Code on a 12-core Snapdragon Elite laptop with 32 GB of RAM. It offers NPU-optimized distillations of DeepSeek's large mixture-of-experts model at 1.5, 7 and 14 billion parameters, which he calls the "squished", "JPEGy" student versions.
• The 14B model works fully offline, but slowly and inaccurately. He switches the laptop to airplane mode mid-answer to show nothing is going to the web, and reports about 40 tokens a second. The model "yaps" through its thinking and gets details about him wrong, so he warns against using it as a fact machine. The 1.5B model runs much faster and loads the machine far less, at a cost in accuracy and breadth.
• Loading the big model took trial and error. The 14B model first failed to load, which he blamed on having too many models in memory, so he killed the toolkit's inference service agent; he later found the toolkit's own "Unload" option. He notes it also offers a playground, a prompt builder, local fine-tuning and free cloud models through GitHub Models, none of which he tried.

2025-03-10 · open the original ↗

• DeepSeek erased the frontier labs' moat, not merely their benchmark lead. Burkov calls it probably the most important event for language models since ChatGPT: a state-of-the-art model trained for what he puts at 5 to 10 million rather than the 20 to 50 million previously assumed, with a published step-by-step recipe that several independent R1-Zero and R1 implementations have already reproduced.
• It is also far cheaper to run. Burkov compares OpenAI's o1 at $60 per million output tokens with DeepSeek's $2, roughly a 30-fold reduction, so that anyone, "your grandmother", can now run state-of-the-art AI.
• Automated verification replaced costly expert-written training data. Burkov argues OpenAI and Anthropic invested heavily, by his estimate billions, in subject-matter experts writing problem-solving dialogues, whereas R1 used binary rewards from checks: does the code compile or produce the right output, does the math answer equal the known value, does the logic puzzle name the right culprit. Hiding a class in a GitHub repo and rewarding a compiling rewrite is one way he says close to a million examples can be generated with no experts.
• Open-weight models are not open source without their data. Burkov's stricter definition, unlike Yann LeCun's, says a model is like a binary: reproducing it requires the training data. The few models released with full datasets are no longer competitive, since today's pre-training uses roughly 15 to 20 trillion tokens versus perhaps four or five trillion in open datasets.

2025-03-10 · open the original ↗

• Colour-code your calendar by how each meeting will feel. Hanselman uses Outlook categories: green for positive personal time such as lunches, red for meetings with bosses or VPs, orange for presentations and podcasts as "approaching stress", pinks and purples for one-on-ones, and blue for general business. He argues a single-colour calendar is like playing Tetris with one colour.
• Grey focus blocks protect time without looking like appointments. He colours focus blocks grey, the same as the calendar background, so the time is held against bookings but doesn't read as a meeting.

2025-03-09 · open the original ↗

• To eliminate hallucination, don't let the LLM generate answers. Burkov says RAG cuts hallucination sharply, by his rough figures from perhaps 20–30% of answers to 1–2%, but not to zero. For his company's closed-domain labor-market product, the team instead fills predefined templates ("the average salary for a nurse in San Francisco is…") with numbers pulled straight from internal APIs.
• Showing the system's interpretation makes remaining errors visible. The LLM only parses the user's string into structured fields, which are normalized to internal taxonomies (e.g. "JS" to JavaScript) and displayed as labels before any number appears. Errors can still come from interpretation or imperfect scraped data, but Burkov says the chatbot's answer matches what the conventional UI would show; the LLM is used "to understand but not to inform".
• Bolted-on uncertainty warnings can be overcautious. The host reports that Claude now often says a topic is niche and should be double-checked, including on questions he doesn't consider niche, which he takes as a layer on top of the core model that still needs tuning.

2025-03-09 · open the original ↗

• Knowing a language's design culture helps you trust it. Torgersen, lead designer of C# at Microsoft (visiting in 1999, joining in 2005), says users choosing a language should look at how much it evolves, how open it is and its track record on developers' problems, as with any vendor.
• C# must keep evolving or become legacy, without ever breaking code. He calls the team "on the aggressive side" of language evolution, adding large features when warranted even though nothing can be removed. With billions of lines of C# in use, much of each feature's work is fitting it in without breaking customers or clashing with existing features, while trying not to follow fads.
• Designing in the open is overwhelmingly positive, despite visible heartbreak. He calls the change "98% positive". Ideas arrive mainly as user-reported problems on the language-design GitHub repo and accumulate until the signal justifies investment. Users watch features get prototyped then shelved, but he says openness makes the conversation human-to-human; Giard's version is that it is easy to hate Microsoft but hard to hate Mads.
• Competing languages are healthy competition and a source of ideas. He would only stop evolving C# if it "can't turn this next corner"; so far it has, including trimming ceremony in response to newer languages. Borrowing, such as index and range support after Python's easy indexing, "borders on stealing but we all do it".
• His proudest contribution is async/await. He drove the feature to end error-prone asynchronous programming, and is most proud that many other languages adopted it.

2025-03-08 · open the original ↗

• AGI is possible but as speculative as teleportation. Burkov declines to give a timeline, saying only that since evolution produced human minds, something analogous might be built.
• Infinite-horizon planning is what LLMs, and other animals, lack. He contrasts crows that use cars to crack nuts and apes that keep a stick overnight but discard it after three or four days without bananas, with humans who plan for generations, as with climate change. LLMs are reactive: an "agent's" agency comes from its system prompt, not itself, so it can be violated unpredictably. The host adds his own hypothesis that AGI may need brain-like modules, such as a prefrontal-cortex analogue, rather than a scaled-up single architecture.
• LLMs cannot tell what they don't know. Burkov asked OpenAI's o3-mini whether his new book, The Hundred-Page Language Models Book, released after its training cutoff, is good; its visible chain of thought admitted knowing nothing about it, yet the answer confidently said experts and readers praise it. He says a Qwen model fine-tuned to answer "I don't know" about obscure people denied knowing him, then named him as author of his earlier book, showing such fixes are patches, not self-knowledge.

2025-03-07 · open the original ↗

• dbt's real value is documented, agreed business logic. In a clip from episode 861, a former pro athlete turned data engineer says dbt let her team embed business logic agreed with stakeholders, with field definitions (dbt docs, now auto-generated, plus reusable doc blocks) that force agreement on what terms like "active customer" mean and help new engineers. She describes Sigma, sitting on Snowflake, as easing adoption for Excel-minded PMs.
• LLM prompts need a compiler the way HTML needed React. Gupta, CEO of Boundary (episode 859), says BAML ("basic-ass machine learning") was inspired by React catching an intern's missing closing div that could break a news feed. Because every codebase will hold more prompts each year, he argues developers consuming models need type-safe tooling with instant error feedback, as TypeScript gave JavaScript.
• A tabular foundation model is spreading across scientific applications. Hutter (episode 863) and the host note v1 was already used in a Science paper on chemical proteomics, and an "awesome TabPFN" repo Hutter had just created lists about a dozen applications, including insurance, fault classification, finance, wildfire propagation and algal blooms. Version 2 handles more data types, missing data, outliers and larger datasets, and can be fine-tuned for domains its authors know nothing about.
• A mandatory paid discovery phase made Pandata scalable. Al-Dhubaib (episode 865) says returning clients valued plain speaking and a 30-day to six-week "discovery and design" engagement, so he made it compulsory, priced then at 12,000 and now 50,000. The pipeline briefly dried up, one of three times Pandata's bank account fell below a month's expenses, but he then landed three of his biggest clients, two of whom stayed about six years until the exit; his lesson is the art of saying no.

2025-03-07 · open the original ↗

• A migration is far harder than copying files. Beaudoin-Allaire, a principal software engineer on ShareGate at Workleap, says you don't own the data or know what you'll find: a thousand documents or a million, odd settings, and sources and destinations with different capabilities. Large files must be streamed from source to destination rather than held in memory, and Microsoft 365 now has chunked upload sessions for them. He recalls the file-size limit growing from 2 GB to perhaps 50 GB or more.
• Microsoft Graph simplified things, but legacy APIs are still needed. Graph gives one way to authenticate and consent, documented calls, and a dependable SDK, unlike older SOAP and CSOM calls that were barely documented. It still has blind spots, so ShareGate keeps an ordered fallback list that includes SOAP calls and exposed SharePoint DLLs, while on-premises versions from 2003 onward need their own APIs. Hanselman calls Graph one of the smartest things Microsoft 365 ever did.
• Every extra call is multiplied a millionfold. Adding one call to the ten or so needed per file version adds up across a million files and can trigger throttling, which he likens to a car braking for no reason and slowing all the traffic behind it. Limits differ by service, shared in SharePoint and per mailbox in Exchange, so work is parallelized to match, and the software is built around fault tolerance in the spirit of chaos engineering.
• ShareGate exposes a performance slider rather than adapting automatically. Beaudoin-Allaire says they cannot tell how close a tenant is to its limits, and must handle both graceful 429 "too many requests" responses and 503s, the service saying it is "on fire". The slider predates cloud rate limiting: on-premises servers could crash under load before signalling anything. Most of the time the network, not the customer's machine, is the bottleneck.
• Metadata that doesn't map gets an approximation and a warning. Permission levels differ between Google and Microsoft, so ShareGate maps them to read and write equivalents and lets customers remap them. Where something cannot be supported, it says so in documentation or warns during the copy. Google Drive files appearing in several places through links were copied to one place, with a warning.

2025-03-07 · open the original ↗

• A blog platform must make leaving easy and make content last. Reece, founder and CEO of Micro.blog, says the service archives pages you link to, keeps a paid blog online unless you ask otherwise, can ping the Internet Archive on every post, and mirrors posts and photos to GitHub about once a week. He says the platform would be "hypocrites" to preach the open web while trapping content, which Hanselman ties to the link rot he finds on working actors' expired sites.
• Micro.blog deliberately stays small. Reece argues that very large, ad-based platforms develop bad incentives and moderation problems, while small fediverse communities can manage themselves. He says tens of thousands of users is fine, and that at 100,000 they would probably have to stop taking new users. He took Hanselman's word "boutique" for it.
• Discovery is hand-curated and shows no follower counts. Reece says trending hashtags let communities clash and invite hateful piggybacking, so a human curates the Discover page from recent posts. An early "people to follow" list was dropped because it snowballed popularity, as Hanselman recalls Twitter's suggested-user list doing. Reece concedes the balance may lean too far toward making people hard to find, and is open to linking distributed starter packs.
• POSSE concedes that not everyone reads RSS. Publishing on your own site and syndicating elsewhere lets you keep control while still reaching people on social networks. Reece says he added cross-posting first for Twitter, because existing tools added unwanted links and he wanted short posts to read like tweets. It now supports eight or nine services, each coded to its API, though platforms such as Instagram lock down posting.
• A $1-a-month plan is an experiment made possible by static hosting. Micro.one launched a week before the recording with domain, photo and podcast hosting, alongside the $5 standard plan, which Reece says will always be $5, and a $10 premium plan. A credit card is required, mainly to deter spammers and AI-generated content farms, and unlimited photo storage is his main cost worry. Blogs are published with Hugo as static HTML served separately from the platform, so they stay up during an outage.

2025-03-07 · open the original ↗

• Visual Studio on Windows has a built-in remote Linux file explorer. Hanselman shows that adding the "Remote File Explorer for Linux" component, about 30 MB under the Linux and embedded development workload in the Visual Studio Installer, puts a Remote File Explorer on the View menu. With it he browses his PiDP-11, a PDP-11 replica kit running on a Raspberry Pi, opens its start-up shell script, has Copilot explain it line by line, edits it with Copilot completions, and saves it back to the Linux machine.
• Visual Studio's terminal can SSH straight into the remote machine. The terminal's connection list includes his WSL distributions and the remote connection, so he opens a shell on the PiDP-11 without any setup and boots Unix V7 on it. In the same terminal he also runs YEdit and htop, which he takes as a sign of how much the team cares.

2025-03-06 · open the original ↗

• The Looneys came to games from NASA engineering. They met at NASA in 1986: Kristin designed VLSI chips starting as a co-op student in 1985, and Andy wrote telemetry software, a file-transfer tool for a real-time OS, and later code for a Hubble coprocessor. Andy, whose father helped found Goddard in 1959, stayed eight years; Kristin, a little longer.
• Fluxx exists because they could not afford pyramid molds. Custom injection molds would have cost $10,000 or more, so Kristin asked for a card game; Andy wrote the memo describing it on 24 July 1996. They say it has run 30 years, sold four and a half million copies, spans about 45–46 versions, and is what "makes us all of our money".
• Andy treats game rules as software run on humans. Each card is like code that supersedes others, play starts with "draw one, play one", and themed versions forced new card types: creepers for Zombie Fluxx, surprises with in-turn and out-of-turn clauses so every card stays playable, the secret in the upcoming Mystery Fluxx. Word Fluxx has 51 keepers against the usual 22, and Solo Fluxx, cracked after 20 years, meant retrofitting every edition to one to six players.
• Luck is deliberate, so anyone can win until the end. They say the balance lets geeks and their non-gamer families play together, at the cost of hardcore gamers' dislike. They also tell fans loudly when only packaging changes, as with Monty Python Fluxx's licensor-required box, rather than manufacture collectible versions.
• The pyramids are a game system rather than one game. Pyramid Arcade bundles 22 games: 20 by Andy, Kristin's Volcano, and Homeworlds by Andy's childhood friend John Cooper, a no-luck strategy game Andy says rivals chess, with a medal for beating him awarded 21 times. Fans have made hundreds of pyramid games, foreign publishers license translations (Fluxx is in at least a dozen languages), and several titles play online at Board Game Arena.

2025-03-06 · open the original ↗

• Executives' agent hype reflects their shareholders' interests. Against the host's list of boosters, Burkov says Nvidia's CEO talks up agents because it implies buying more GPUs, and that Salesforce once declared traditional software dead in favour of SaaS, which did not happen, and now says the same of SaaS. He distinguishes Andrej Karpathy as careful to describe possibilities rather than predict them.
• A single agent is a rebranded LLM, useful only in-distribution. Burkov's best case for agents is gathering and summarizing web information, because web data is closest to what LLMs were trained on; he questions how big that market is, likening some of it to a cron job running Google searches. Dropped into a company's intranet of legacy systems they have never seen, he expects agents to "break their teeth" quickly.
• LLM-based multi-agent systems are effectively undebuggable. Burkov, whose PhD was in multi-agent systems, explains that distributed systems are already hard to debug because you cannot pause all asynchronous actors at once, and LLM agents add black-box internals you cannot step through. Reliable multi-agent systems need full control of each actor and inspectable communication protocols.
• Agents need human-level trustworthiness before being combined. Burkov argues that even human systems guard against unreliability with redundancy, such as two keys turned far apart, and that today's agents are nowhere near as reliable as a person, so multi-agent systems built on them are "a recipe for disaster". He says no one has a clear idea how to reach a trustworthy AGI.

2025-03-05 · open the original ↗

• Systems thinking matters more as AI does the coding. Hanselman relays a systems architect's argument, from what he calls his most popular Hanselminutes episode, that the learn-to-code movement taught for-loops but not the systems code lives in. Her example is a cheap Amazon package that arrives next day with nobody thinking about the oil, freight, labour and macroeconomics behind it. Mark agrees and equates systems thinking with architectural thinking, saying that with AI writing so much code, the value lies in the thinking on top of it.
• Sketch the rough shape, then sculpt iteratively. Mark says he starts a Sysinternals feature from a broad structure and reworks it as he goes, likening coding to sculpture where the chisel strikes can't be predicted. Hanselman says his own systems thinking leaves him stuck among nine ways to do something. In brownfield code, both say you should sculpt in the previous sculptor's style rather than bolt things on.
• Nobody holds today's systems whole, so learn just enough to decide. Mark relays Scott Guthrie being asked who still has Azure in their head; the answer was nobody, and Mark says he knew a far larger share of Azure 10 to 15 years ago. As a senior executive reviewing experts' work, he asks how much he needs to know to make a call, looks for risk areas from experience, and asks questions where there are gaps.
• Mark sees "no line of sight" to AI replacing programmers. His first reason is specification: you never fully specify what you want, the AI's fixes break other things, and you end up owning code you didn't write, which leaves a non-coder "dead in the water". Hanselman's example is a site built with his 17-year-old, where the AI made red ones-and-zeros "Matrix rain" from 256 hard-coded spans that overheated his laptop. Mark says over 95% of his own research code is AI-generated, which he can manage only because he understands it.
• AI is stuck on old APIs and thin documentation. Mark says a model generated code against an older OpenAI SDK because it had seen more examples of it, and that generated reference docs lack the context a human or an AI would need to use a new API. His conclusion is that AI accelerates experts but limits beginners to simple projects, so young people still have to learn systems and how to program.

2025-03-05 · open the original ↗

• TalentNeuron turns scraped job postings into workforce insights with NLP. The host cites Burkov as having previously said the company makes 95 million daily predictions from 700 million job-posting data points. Burkov describes some 35,000–40,000 rule-based robots downloading postings, then models that extract and normalize skills, telling "excel in Word" from the skill Excel, and predicting industries with BERT and RoBERTa since around 2018–2019.
• Multilingual models removed the need to label data per language. Where the team once built separate salary extractors for about 25 languages, Burkov says they now reuse labeled data from other languages to cover new countries, a capability he says he would have called impossible ten years ago.
• LLMs now build about half of the company's scrapers automatically. Given a corporate home URL, the system finds the careers section and writes extraction rules for titles, locations, dates and descriptions; the remaining harder half is still done by engineers.
• Think hard before putting an LLM in front of users. Burkov tells decision-makers to think "30 times", warning that LLMs work in-distribution but businesses cannot know what is in-distribution because training data is hidden, even for open-weight models. Systems that pass developers' tests can fail arbitrarily on real production data, with no way to detect the failure or know whether more fine-tuning examples will fix it, costing reputation and angry customers.

2025-03-04 · open the original ↗

• LLMs are for prototyping; production still needs trained, controlled models. Burkov, ML lead at TalentNeuron, warns that businesses cannot know whether their problem is in-distribution because training data is hidden, so systems that pass tests fail unpredictably on real users. His four-person team uses instructed LLMs as placeholders at the many decision points in a pipeline, then replaces the weakest links with fine-tuned or purpose-built classifiers, because errors such as 15% at several steps compound toward zero end-to-end success.
• Agents won't work in real life, and multi-agent systems can't be debugged. Burkov, whose PhD was in multi-agent systems, calls a single agent a rebranded LLM with the same out-of-distribution problems, useful mainly for web information gathering. Asynchronous black-box agents cannot be paused and stepped through, so he calls building multi-agent systems on them "a recipe for disaster" until agents are as trustworthy as humans. He attributes CEO hype from Nvidia and Salesforce to shareholder interest.
• LLMs lack infinite-horizon planning and a sense of what they don't know. Burkov likens AGI's prospects to teleportation and declines a timeline. Humans plan for generations where apes discard a tool after three or four days, while LLM agency comes only from the system prompt. His test of OpenAI's o3-mini on his new book, released after its cutoff, produced a reasoning trace admitting ignorance but a confident answer saying experts praise it.
• Closed-domain chatbots avoid hallucination by keeping LLMs from writing answers. At TalentNeuron the LLM only parses queries into structured fields, which are normalized and shown to the user, while numbers come from internal APIs into fixed templates. Burkov contrasts this with RAG, which by his rough figures cuts hallucination from perhaps 20–30% to 1–2% but not to zero.
• DeepSeek removed the frontier labs' moat. Burkov says it trained a state-of-the-art model for 5 to 10 million instead of 20 to 50 million, published a recipe others have reproduced, and charges $2 per million output tokens against o1's $60. Its R1 training used automatically verified rewards (code compiles, math matches) rather than expensive expert-written data. He adds that open-weight models are not open source without their training data.

2025-03-03 · open the original ↗

• Sell to the buyer where your business creates the most value. Al-Dhubaib says firms had long approached Pandata because buying a data science team is easier than building one. When the head of Further (then Search Discovery), a fellow Case Western alumnus seeking to move into cloud, AI and advanced ML, offered to buy within five minutes of meeting, he first asked about core values to judge whether his clients and team would thrive there.
• Upfront compatibility enabled an unusually fast deal. Al-Dhubaib found the two firms shared problems, attitudes and pace despite differently named values; diligence and paperwork ran from January to March of last year, about three months. He attributes the horror stories he hears about other exits to skipping that matchmaking, partly because it is hard to say no to good-sounding offers.

2025-03-02 · open the original ↗

• Evaluation design and LLM operations are today's scarce data science skills. Al-Dhubaib says what "data scientist" means has changed drastically over ten years; he now seeks people who can design statistically sound evaluation and quality control for unstructured, multimodal data without leaking what they predict into the data, and ML engineers experienced in deploying language models and diagnosing drift or odd behaviour, which he calls AI Ops.
• The evergreen traits are soft skills, not technical ones. These are explaining plainly, a drive to learn, and comfort with ill-formed problems. Whatever the technical skills of the moment, Al-Dhubaib says these core values from Pandata are what he consistently looks for and recommends data scientists cultivate to stand out.

2025-03-01 · open the original ↗

• Without an upfront value hypothesis, AI projects fail on ROI. Al-Dhubaib says data scientists get enamored with model accuracy instead of asking which decisions the model influences, what a successful intervention is worth, and how often the decision will actually change.
• A risk score is useless unless it tells providers what to do. In a hospital project on solid-tumor cancer patients, whose national readmission rate he puts at about 25%, clinicians responded to risk alerts with "now what?". The team added interpretability showing why a patient was at risk, such as medication, comorbidity or a zip code with unreliable transport, pointing to interventions like social assistance.
• Stress-test the value estimate with the client before investing. In the readmission example the team assumed it could identify 80% of at-risk patients but only 20% would receive an intervention, valued the project on that basis, and asked whether it was still worth the investment, revisiting the hypothesis as models were validated.

2025-02-28 · open the original ↗

• Warp breaks terminal conventions to behave like a notebook. Lloyd says the mouse and text selection work as in a normal editor, and each command is married to its output as a "block" that can be shared, filtered or attached as AI context. He accepts some compatibility cost for a native GUI over "terminal purity": "rules are made up."
• Agent mode drives the terminal, asking approval by default. An on-device classifier decides whether input is a command or English; agent mode can run commands such as git status to gather its own context, as in his merge-conflict example. Users can opt into LLM-judged read-only commands or allow and deny lists; models come from OpenAI, Anthropic and DeepSeek (R1), with Claude 3.5 Sonnet the default "as of today".
• AI, not collaboration, is what now pays for a terminal. Asked about a reported $50 million raise, Lloyd says free usage made the early years stressful, and Copilot and Cursor changed developers' expectation to paying for AI's marginal cost; he says revenue is "growing pretty well". His stated aim is "the world's best place where developers can build software", with the terminal as the starting point, and Warp Drive extends shared team commands into context for its AI.
• Windows came last because Warp went deep on Mac first. Founded June 2020 and written in Rust, Warp shipped on Mac, then Linux; Windows was the top feature request, serious work began about a year before recording, and the Windows build launched the day before. Lloyd says he would now do Windows before Linux and concedes Windows Terminal is more polished in places; the episode is sponsored by Warp, and Hanselman says he spent hours with the team on its Windows experience.
• Offline-first, with AI and telemetry under user control. Everything except AI and real-time collaboration works offline, one toggle removes all AI, and local models via Ollama are being looked at but not built. Smart completions with flag explanations come from a local database built on open-source completion specs.

2025-02-28 · open the original ↗

• AI services are Cognitive Services under a new name. Giard opens the project created in the previous episode at ai.azure.com and finds AI services in its left-hand menu, in four categories: speech, language and translator, vision and document, and content safety. Vision and document has three tabs: document, face and image. Every service he shows returns JSON over REST calls, and most have an SDK, sample code in Python, C# or JavaScript, documentation and a quick start.
• Document templates parse receipts into structured fields. The document tab has templates for invoices, receipts, ID documents and credit cards, plus a generic optical character recognition "read". Run on a sample receipt, it found the address and split it into house number, road and postal code, and listed the line items. It also parsed a hotel receipt of Giard's own.
• Face detection returns mask status and facial landmarks as JSON. For each face it reports whether a mask is worn and whether it covers the nose and mouth, the face rectangle, and landmark positions such as the pupils, nose tip, mouth ends and eyebrows. A photo with three faces came back as an array of three. Giard notes that the bounding box and landmark dots are drawn by the sample app's JavaScript; the service itself returns only the JSON.
• Object detection and captioning label image contents with confidence scores. Object detection returns each object with a bounding box and a confidence, e.g. 98% for a skateboard and 95.5% for a person in a sample, and 57% for a bird in Giard's own photo of a guitarist. Image captioning gave one caption for his bird photo, "a green bird on a branch", which he calls "pretty darn accurate". Dense captioning adds captions for different regions of the same photo.
• OCR returns text by line and word, each with its bounding box. On an image of a poem it split the text into lines, and each line into words, with the X, Y coordinates of the four corners of each box. Common tag extraction returns a list of tags with relevance scores, e.g. 96% for a picture frame and 95.5% for a gallery in a photo of a painter in an art gallery.

2025-02-28 · open the original ↗

• Investors are pricing de-extinction on its side technologies, not revenue. The host reports that Colossal Biosciences raised a $200 million Series C at a valuation over $10 billion, nearly sevenfold its valuation a couple of years earlier, with no revenue yet and its main projects not projected to complete until 2028.
• The method is genome comparison, CRISPR editing and a surrogate. As the host describes it, Colossal maps the extinct species' genome, compares it with a close living relative (the Asian elephant for the mammoth), edits the relative's cells with CRISPR to match, and implants the resulting embryo in a surrogate; it has also built artificial-womb technology the host says could apply to human fertility treatment.
• Colossal makes money three ways while chasing its core mission. The host lists spin-offs (one using microorganisms to break plastic waste into organic matter, and Form Bio, a computational biology platform for gene therapies, both with millions in further funding), conservation work with governments on species such as the northern white rhinoceros, of which two females remain, and an emerging biodiversity-credits market modelled on carbon credits.
• De-extinction needs good DNA, a living surrogate and a surviving habitat. On those conditions the host rules out dinosaurs; he names the dodo and Tasmanian tiger as active Colossal projects, the passenger pigeon as a Revive & Restore project, and the Caribbean monk seal, quagga, great auk and heath hen as other candidates.
• The host's lesson is that impossible goals throw off valuable byproducts. His opinion is that each of Colossal's indirect applications could eventually generate billions of dollars in revenue on its own, and he urges listeners to use AI tools such as OpenAI's and Google's Deep Research features, new "as of the past few weeks", to dream up ambitious goals.

2025-02-27 · open the original ↗

• Repeat clients, not interest, revealed Pandata's product-market fit. Al-Dhubaib says that because he bootstrapped, paying customers were his only source of growth; when he asked one of the few returning clients why it came back, the answers were his firm's plain, jargon-free approach and an upfront scoping phase it had stumbled into.
• A mandatory paid discovery phase screens out the wrong clients. He made a 30-day to six-week "discovery and design" engagement compulsory, priced at $12,000 at the time and $50,000 now, to find "where the skeletons are" and the unknown unknowns; it filtered out clients who weren't serious and those who didn't accept the data science process and would end up disappointed.
• The filter nearly broke the firm before it paid off. Al-Dhubaib says his pipeline dried up and the bank account fell below a month of expenses, one of three such times, but in that same period he landed three of his biggest clients, two of which stayed for about six years until Pandata's exit.
• Saying no to small, low-value work is what lets a consultancy scale. He calls ruling clients out a counterintuitive lesson that frees time for higher-value work, and says the acquirer's due diligence flagged about three points in the financials that matched the moments he had regretted too.

2025-02-26 · open the original ↗

• A hub is the top-level resource; projects are containers inside it. Giard describes Azure AI Foundry as a unified platform to explore, build, test and deploy AI applications. The hub is its primary top-level Azure resource, a project is a container for your AI work, and one hub can hold several projects.
• Creating a project also creates or selects its hub. At ai.azure.com, the Create Project button asks for a project name and a hub, either an existing one or a new one. A new hub name must be unique and between 3 and 33 characters.
• The hub step provisions the supporting Azure resources. Giard sets the subscription and creates a new resource group, keeps East US as the location, creates a new Azure AI service or OpenAI service, and also creates an Azure AI Search resource, which is optional. Validation checks for problems such as a region where the services are not available, and the Create button is enabled only when there are none.
• Provisioning took about three minutes in the demo. Status icons show each resource, including the project, a storage account and the AI Search account, as queued, in progress or created. The finished project shows an API key and a left-hand menu with model catalog, playgrounds and AI services. Giard notes that a few of the items are in preview.

2025-02-26 · open the original ↗

• The failed startup's free research was the real product. Al-Dhubaib says his first venture, built out of computational neuroscience work at Case Western, raised money, entered an accelerator and collapsed in about a year and a half; hospital researchers he had been doing free pilot research for then told him they had depended on it, and would pay for it.
• Pandata was founded on a regional data science talent gap. At its start, at the end of 2015 and start of 2016, he says Northeast Ohio had fewer than 150 data scientists, about a third of them at IBM, the Cleveland Clinic or Progressive Insurance, while other large local enterprises had one or two.
• The firm survived by moving to whatever was still hard. He says it shifted from analytics and business intelligence to machine learning, then to machine learning in heavily regulated environments as each layer became commoditized; after about eight bootstrapped years it was acquired by Further, where he is now head of AI and data science.
• "Make it boring" became a core value because clients who understand stay. Al-Dhubaib says many data scientists of the time liked being opaque, while he found clients who grasped the work hired the firm and stayed longer, so he explains complex mathematics with cats, dogs and memes.

2025-02-25 · open the original ↗

• The UCG-Fiber lets any copper port serve as either WAN. It has four 2.5 Gbps RJ45 ports, one 10 Gbps RJ45 port and two 10 Gbps SFP+ cages, and any of the 2.5 or 10 Gbps ports can be set as primary or secondary WAN. It runs UniFi OS with the full app suite (Network, Protect, Access, Talk), and an M.2 NVMe slot provides Protect storage.
• The host got full LAN wire speed in testing. He set WAN on a 2.5 Gbps port and ran a 10 Gbps speed test across the LAN. The figure of just over 5 Gbps WAN-to-LAN with IDS/IPS enabled is stated, not shown tested. Base price is $279 with no storage; the drive tray is a separate accessory, and 1 TB and 2 TB NVMe versions are also offered.

2025-02-25 · open the original ↗

• The UDR 7 is an all-in-one gateway needing no separate controller. It runs UniFi OS and UniFi Network on board, so no Cloud Key or separate UniFi Network controller is needed.
• The UDR 7 adds an SFP+ port and dual WAN. Around the back are four 2.5 Gbps RJ45 ports, one with PoE output for an access point, VoIP phone or camera, plus a 10 Gbps SFP+ slot; internet can run off RJ45 or SFP+, and dual WAN is supported.

2025-02-25 · open the original ↗

• The UniFi Express 7 upgrades the original Express to Wi-Fi 7. The UX7 is physically larger than its Wi-Fi 6 predecessor and has a 2.5 Gbps RJ45 LAN port, a 10 Gbps RJ45 WAN port and 2x2 Wi-Fi 7. In the box are a USB power brick with cord and a braided patch cable.
• Like its predecessor, the UX7 runs UniFi Network only. The host found Network was the only application offered, so it cannot run Protect, Access or Talk. Setup was as simple as other UniFi gateways, and he updated it to UniFi Network 9.0.1 without issue. As of recording he did not know the MSRP or how many UniFi devices it can manage.

2025-02-25 · open the original ↗

• The UCG-Fiber is the host's pick for multi-gig internet connections. It has a quad-core ARM Cortex-A73 CPU and 3 GB of DDR4 RAM, and Ubiquiti claims over 5 Gbps of throughput with IDS/IPS enabled. The host did not run a multi-gig WAN-to-LAN test himself and asked viewers to post theirs.
• Storage costs extra on the base model. The $279 base model has no NVMe drive; adding your own requires the $19 M.2 drive tray. Versions with 1 TB and 2 TB pre-installed were launching, but the host had no pricing on them at recording. Like the UCG-Max, the NVMe slot is for Protect surveillance storage.
• Any of the seven ports can be primary or secondary WAN. There are four 2.5 Gbps RJ45 ports (one with PoE+), a 10 Gbps RJ45 WAN port and two 10 Gbps SFP+ cages, one WAN and one LAN by default. The host put WAN on a 2.5 Gbps port and disabled secondary WAN, freeing all three 10 Gbps ports for LAN.
• The host measured full LAN wire speed in his lab. He got 10 Gbps throughput running OpenSpeedTest on a Minisforum MS-01 over a DAC on an SFP+ port and testing from a Mac Mini on the 10 Gbps RJ45 port. He suggests 2.5 Gbps WAN users can use the three 10 Gbps ports for a core-switch uplink, a NAS and a workstation.
• He expects price and mounting to be the only complaints. He struggled to find valid criticisms. It is desktop-only, powered by a separate 54 V DC brick, with a template for the rubber feet. He discloses that his purchase link is an affiliate link that pays him a commission without changing the buyer's price.

2025-02-25 · open the original ↗

• The UDR 7 keeps the old form factor with faster ports. The old UDR had a single 1 Gbps RJ45 WAN; the UDR 7 has four 2.5 Gbps RJ45 ports (one with PoE out) and a 10 Gbps SFP+ slot, adds a Wi-Fi 7 access point on top, and runs UniFi OS and Network on board with no separate controller.
• Either WAN can be assigned to any of three ports. The choices are ports 3 and 4 (2.5 Gbps RJ45) and port 5 (the SFP+). The host ran Comcast Business cable on port 4 and T-Mobile Home Internet on port 3 as dual WAN, and reassigned the SFP+ (secondary WAN by default) to LAN for a 10 Gbps DAC link to an OpenSpeedTest server.
• Wired tests reached full wire speed on every client. A 2.5 Gbps-connected Minisforum server hit full 2.5 Gbps across the UDR 7 to the 10 Gbps server, and a 1 Gbps laptop got essentially full gigabit. The setup speed test on his sub-gigabit cable line showed 932 Mbps down and 237 Mbps up.
• Wi-Fi 7 tests needed channel widths maxed out first. After forgetting this initially, he set 240 MHz on 5 GHz and 320 MHz on 6 GHz. A 2x2 Wi-Fi 6 iPhone 13 Pro Max reached roughly 770–790 Mbps, and a 2x2 Wi-Fi 7 Google Pixel 8 roughly 1.7–1.9 Gbps, though the captions garble the exact figures.
• He recommends it as an all-in-one for newcomers and small sites. He sees it suiting parents' homes or small businesses he administers, and it can still expand with access points, Protect or UniFi Talk phones. Its microSD slot handles Protect for one or two cameras; beyond that he points to a UNVR. Price was unknown at recording.

2025-02-25 · open the original ↗

• A mandatory paid discovery phase built Pandata's durable client base. Al-Dhubaib made a 30-day to six-week "discovery and design" engagement compulsory, at $12,000 then and 50,000 now, to weed out unserious clients and those who would not accept the data science process; the pipeline dried up and cash fell below a month of expenses, then he landed three of his biggest clients, two of whom stayed about six years until the exit.
• Mid-sized regulated work too small for the big firms was the niche. He says Pandata served hospitals, life sciences, higher education, energy and utilities and some defense work, built insurance and security processes "way more secure than any of our clients", charged a premium, occasionally took projects from McKinsey or Accenture, and that this expertise was a big reason Further acquired it.
• Building users' trust lifted a failed model without changing it. In his telling, an ensemble model for a health-tech client with about 50 million patient lives found 30% more people eligible for government assistance, but its pilot converted 2 in 100 mailings against a baseline of about 10, because staff with bonuses tied to selection distrusted its unfamiliar picks; after training sessions and per-prediction explainability, the same model grew reach by 18%.
• Projects need a value hypothesis about decisions, not model accuracy. For cancer readmissions, with a national rate of about 25%, Al-Dhubaib says providers asked "now what?" until the model explained the drivers of risk, such as poor transport access; he now stress-tests value up front, for example identifying 80% of at-risk patients with a realistic 20% intervention rate.
• An acquisition works when the buyer's values and clients fit. He says he asked Further's CEO about core values within minutes of being offered a deal, to judge whether clients and staff would carry over, and diligence closed in about three months between January and March of the previous year; his hiring values, he says, are approachability, winning together, cultivating trust, pursuing growth and taming uncertainty.

2025-02-24 · open the original ↗

• A university group could not build AutoML tools for the masses. Hutter says that after 10 to 12 years co-leading the AutoML community, whose best-known tool, Auto-sklearn, wraps scikit-learn to choose methods, preprocessing and hyperparameters, a lab with a few research engineers only ever built tools fit for researchers; reaching a broad user base, especially with TabPFN, needs a commercial entity with strong engineers.
• Prior Labs builds products while the university keeps the moonshots. Hutter, co-founder and CEO of Prior Labs, says he will keep an academic co-affiliation and research tabular-data questions there, such as how the learned algorithm works, how it shifts with the priors and what its failure modes are; PhD students gain engineers who can turn their papers into products.
• Prior Labs is hiring broadly and putting community before sales. At recording, Hutter says a funding announcement was due two days later; the company was hiring AI scientists, ML engineers, backend engineers and community staff, with sales left for later.

2025-02-23 · open the original ↗

• Even TabPFN version 1 found real scientific and industrial uses. The host cites a Science paper on large-scale chemoproteomics for ligand discovery that used TabPFN in its inference, and an "awesome TabPFN" GitHub repo listing about a dozen applications across health insurance, factory fault classification, finance, wildfire propagation and biology.
• A generic method lets non-experts fine-tune into specialist domains. Hutter says he knows nothing about chemical proteomics but could fine-tune TabPFN on that field's data to beat the general model, and he names algal-bloom prediction and many medical applications; he expects finance uses exist but go unpublished.
• Prior Labs wants the open-source community to carry adoption. Hutter says he created the applications repo the day of recording and invites pull requests; within a couple of days of the Nature paper, an open-source interpretability repo had integrated TabPFN, and the host notes version 2 handles missing data, outliers and larger datasets.

2025-02-22 · open the original ↗

• TabPFN's prior is a generator of random structural causal models. Hutter credits first author Noah Hollmann with a method that samples graphs in which features and the target are nodes with unknown causal direction; real data never reveals the true graph, so the ideal prediction is a Bayesian posterior that weights each candidate model by how well it fits.
• The network learns to approximate that posterior in one forward pass. Hutter says TabPFN never stores the roughly 130 million structural causal models used to train it; it sees only the raw data and has learned to interpolate across all of them, which the host frames as a substitute for the trillions of tokens language models get and tabular data lacks.
• Training only on synthetic data rules out test-set leakage. Hutter says reviewers rejected the first TabPFN v1 submission because its performance looked too good, suspecting tuning on the test set; the team dropped a component that made gradient updates on real datasets, lost almost nothing, and since then never touches real data in training.

2025-02-22 · open the original ↗

• Recognizing only your own commands makes local voice fast and accurate. Hansen built Rhasspy around asking "which of these things did the user say" rather than "what did the user say", using Kaldi with a bespoke language model of a few hundred kilobytes in place of a gigabyte one. He reports sub-second responses on a Raspberry Pi 4 and under a tenth of a second on a desktop, but close options such as 15 and 50 need clearer speech, combinatorics can blow up to millions of sentences, and rejecting out-of-scope speech only recently became robust.
• Local-first means choice, not rejecting the cloud. Hansen disputes Hanselman's "vote of no confidence" framing: components are modular, Nabu Casa's cloud routes speech through Microsoft Azure, and requests the local system cannot handle can fall back to ChatGPT, Claude, Gemini or a local model via Ollama. He cites quadriplegic users who ran Alexa, Google and Siri together for redundancy and wanted something more robust.
• ESPHome hardware is how voice reached Home Assistant. Hansen, who works for Nabu Casa, says its 2023 "year of voice" push extended ESPHome firmware, and the company sells the $59 Home Assistant Voice Preview Edition with an XMOS audio chip, "preview" because the software is still a work in progress. The device keeps only enough audio to detect the wake word, then streams to the user's Home Assistant server, where Whisper suits beefy hardware and an add-on he is building trains itself on the home's device names for a Pi.
• Training data, not models, limits accents and languages. He says wake-word data lacks accented English and female or high-register voices; his Kaldi models use public-domain Mozilla Common Voice data, and custom pronunciations can accommodate accents. For his Piper text-to-speech voices about a dozen volunteers record two to three hours each, Nabu Casa pays for studio time in some languages and releases the data to the public domain, and LibriVox audiobooks train poorly because narration lacks short phrases.
• Cloud assistants are declining as local LLM hardware lags. Hansen says Alexa and Google are getting worse as their makers fail to profit, while local LLM boxes cost two to five thousand dollars and are not plug-and-play with Home Assistant. He wants an in-home AI box, which would allow uses such as a vision model describing a camera snapshot every five seconds that a paid API would make too costly.

2025-02-21 · open the original ↗

• o3-mini high is the top public model on key reasoning benchmarks. The host says that in its high mode, the one with the most inference-time compute of three (low, medium, high), OpenAI's newly released o3-mini beats o1-mini, DeepSeek R1 and the full-size o1 on the AIME math, Codeforces coding and SWE-bench Verified benchmarks.
• o3-mini matches or beats o1 at a fraction of its price. The host puts o1 at $60 per million output tokens and o3-mini at 7% of o1's price on both input and output; he says o3-mini costs about twice as much as R1 on DeepSeek's own cloud in China, but about half as much as R1 run on a US cloud provider.
• The host rates the release below R1 because o3-mini is proprietary. He says R1's open-source release gives more freedom to adapt the model and to run it on any infrastructure; o3-mini is available to free ChatGPT users through the "reason" button, to paid plans as o3-mini high, and through the API.
• OpenAI's own figures put the full o3 far ahead. The host relays OpenAI's not yet independently verified scores: on AIME, 96.7 against 87.3 for o3-mini high and 79.8 for R1; on Codeforces, an Elo over 2,700 against about 2,100 and 2,000; on SWE-bench Verified, almost 72 against about 49; on GPQA, almost 88 against 80. He expects o3's public release soon and takes OpenAI's past figures as fairly reliable.

2025-02-20 · open the original ↗

• The LAN Turtle gives remote shell access to any network it joins. The Hak5 device looks like a USB Ethernet adapter but is a small Linux box; the host configured it to open a key-authenticated reverse SSH tunnel to an Ubuntu server at Vultr, then SSHes to port 2222 on that server to land inside whatever network it is plugged into.
• MAC address whitelisting will not stop it. The device can mimic the MAC address of the host computer it is plugged into via USB. The host calls detecting and preventing such devices a hard question, asks viewers for their approaches, and says a follow-up video on defences is coming.

2025-02-20 · open the original ↗

• Prior-data fitted networks are trained to compute the Bayesian posterior predictive. Hutter confirms that this is the optimization objective of every PFN, so a Transformer is carrying out Bayesian inference.
• PFNs replace hairy MCMC and variational inference with plain supervised learning. Hutter says closed-form posteriors exist only for simple cases such as Bayesian linear regression or Gaussian processes with known hyperparameters; beyond that, MCMC is slow to converge and variational inference carries approximation error, whereas a PFN samples millions of datasets from the prior and learns with cross-entropy loss to predict held-out points, approximating Bayesian inference "to an arbitrarily strong degree".
• A PFN can integrate over architectures, not just weights. Hutter says standard Bayesian neural networks place a posterior over the weights of one fixed architecture; with a PFN you can place a distribution over architectures too and get the posterior predictive in a forward pass, which he calls Bayesian neural architecture search.
• The limitation is that PFNs give no posterior over weights. Hutter says they go straight to the predictive distribution of y given x, so they cannot say which architecture or weights are right, only give the Bayesian average over all of them.

2025-02-20 · open the original ↗

• Explorer's inconsistencies are the price of decades of compatibility. Hanselman calls some of the hate toward Windows 11 File Explorer unfair but shows its rough edges: a flash of light mode in dark mode, and modern menus that open older-looking dialogs. He says Explorer carries COM objects, XAML islands and plug-ins that a from-scratch app can ignore, and that search improvements are underway.
• Files is a polished open-source alternative aimed at aesthetics. He believes files.community is largely one main developer with many contributors; it lists WSL and cloud drives such as Dropbox separately, gives removable drives an eject button, and adds a Git column showing status and the latest commit message. It also enumerates the shell extensions plugged into Explorer under "show more options", and supports tear-off tabs; it is on the Windows Store.
• File Pilot is a tiny from-scratch beta built for speed. The whole app is 2 megabytes, it is free during the beta with a license mentioned for the future, and he guesses it is written in C or possibly Rust. He shows instant Ctrl+T tab spam, a fast filter and an Everything-style global filter, asynchronous folder sizes on hover, middle-click split views, and an open-console command, but warns its default font is "shockingly small" and notes there is no cloud-drive call-out yet.
• The competition is good for Explorer. Hanselman calls the two apps "effectively a vote of no confidence against Explorer" but welcomes them as unapologetically Windows-native apps that give the Explorer team something to aspire to. He also points to Midnight Commander, by Miguel de Icaza, as a Norton Commander-style option.

2025-02-19 · open the original ↗

• Mark wrote ZoomIt for his own Windows internals demos. He says he built it because a Toshiba laptop zoom utility was clunky for the Sysinternals demos he gave with Dave Solomon, and it has evolved for 20 to 25 years. Most of it sits in one C file around a 7,000-line function; it now ships in PowerToys and is open source in the Sysinternals GitHub repo, where Mark says a refactoring PR would be considered.
• ZoomIt is an everyday tool, not just a presenting aid. Mark says he zooms almost daily on fonts he cannot resize, and Hanselman calls it the first thing he installs. Against Windows Magnifier, which they describe as an accessibility tool with a floating toolbar, they stress ZoomIt's keystroke-driven annotation, which Hanselman likens to John Madden drawing on screen; he says people with eyesight impairments and teachers use it too.
• The feature set runs from zoom-and-draw to live zoom and recording. Mark demos Ctrl+1 to freeze and zoom, drawing with shapes, letter-keyed colors, highlighting and blur, typing, Ctrl+2 to annotate unzoomed, and the newer live zoom that annotates a moving screen. Ctrl+6 copies a screenshot or crops one to file, and Ctrl+5 records the screen or a region; there are also break timers and black or white whiteboard screens.
• Fluency in ZoomIt comes from drilled practice, like a kata. Hanselman argues that pointing, talking and drawing at once only looks effortless because it is practiced, and grades it in improvised belts from simple zoom-and-click up to typing while talking. His tips are to point at the target before Ctrl+1 so the zoom lands there, and to avoid nauseating the audience, though he could not demo live drawing because his build lacked it.
• DemoType replays prepared code as if typed live. Hanselman says he pitched it to Mark after the joke site hackertyper, and Mark built it with others' contributions: snippets in a text file, split by start and end markers, are typed out automatically or driven by random keystrokes, and can include Enter and arrow keys to walk through IntelliSense. Hanselman concedes it is not yet in his muscle memory, though he says it improved his South Africa keynote.