Souk's Notes - Unofficial Summaries
Polis Pandit$The Best Form of Generational Wealth Is Not Money
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon Krohn913: LLM Pre-Training and Post-Training 101 — with Julien Launay
synopsis
Julien Launay, co-founder and CEO of Adaptive ML and formerly extreme-scale team lead at Hugging Face and an industrial PhD researcher at LightOn, argues that reinforcement learning in post-training is becoming the main stage of LLM development. In his view it is the cheapest route to frontier-level performance on a given task from a smaller open model, and the main way past the limits of human-written data. He ends on a forecast: superintelligent systems within about five years (10 for bears), arriving unevenly. He expects the transition to be messy and what follows to be, probably, the best time ever.
Pre-training and post-training. Pre-training predicts the next token over "nearly every text in existence," plus images and video, on tens or hundreds of thousands of GPUs. The raw result is unwieldy: asked a question, it may produce 10 similar questions. Post-training sharpens it for use, and its big success has been RL, learning from a good or bad signal rather than from demonstrations. The phases are blending. Pre-training shifts from bulk data toward "higher quality" code, maths and chat, and RL is creeping into pre-training. Launay says some recent models, Grok 4 among them, spent nearly as much on post-training as on pre-training, and he expects post-training to dominate.
Kinds of signal. RLHF from human annotators doesn't scale well. Verifiable or execution feedback, from maths answers, code tests mined from GitHub or game scores as in AlphaGo and Atari, scales readily. AI feedback from judge models covers tasks that can't be verified. To skeptics who fear model-generated data collapses, Launay answers that it can be done badly but usually works. He says agreement between human annotators is far lower than assumed and model-human agreement matches it, and that verification is easier than generation. The host reports finding GPT-4 indistinguishable from his own team's expert raters a year earlier.
Why RL generalizes. Launay's analogy is that supervised fine-tuning is like memorizing a worked problem about spinning black holes and failing once the hole carries a charge, while RL lets the student try and be corrected. RL is mostly online: samples come from the current weights, so they stay in the model's distribution and reshape it gradually. SFT drops in a new distribution and hopes it connects to old knowledge. He concedes pre-training generalizes by sheer diversity, and that post-training is now run at larger scale too. PPO and GRPO share these properties and differ mainly in credit assignment: PPO trains a value model to spread a sparse reward across tokens, GRPO averages over several rollouts.
What Adaptive does. Launay says RL's obstacle is engineering. It mixes inference and training, and needs judges (Adaptive runs five to 10 per pipeline) and environments: SQL databases for text-to-SQL, compilers, tools, VMs for computer use. Adaptive's "RLOps" tooling turns a data scientist's Python recipe into a distributed run, starting from open models such as Llama, Qwen or Kimi. It never trains from scratch, because a pre-trained model makes RL "easy mode." He claims RL always yields the best trade-off between cost and performance. The firm, a bit over 18 months old, sells to large enterprises through its sales team, but synthetic data's unexpected strength makes production data optional, so it plans wider access. He claims an 8B model can reach frontier performance on a task mostly with synthetic data. In one client project, about 70 human critiques and rewrites seeded about 80,000 synthetic conversations of roughly 10 turns each, and with exploration nearly 10 million data points. AT&T, a named client, now runs it independently. He stresses that RL work is about defining and measuring success: "if you can measure it, you can optimize it."
Data and compute. Launay sees "a bit of truth" in both Ilya Sutskever's claim that web data is exhausted and the CEOs who see no end to scaling. Data can be cleaned and reordered, but there are only so many human words. In Richard Sutton's terms, the way forward is learning from experience, in billions of virtual environments and eventually real-world experiments. Compute is the bottleneck now, but, having worked on LightOn's photonic chip, he is bearish on new paradigms: an Nvidia chip embodies a decade-long supply chain no rival can quickly reproduce, and GPUs are already specializing for transformers.
Superintelligence. He expects capability to be "spiky." Maths can be solved in a closed box with formal verification, while drug and vaccine development stay bound by experiments. Scaling AI-controlled wet labs is the next bottleneck he names. Some fields may advance more in a year than in all of history while others barely move. That is the nuance he adds to fast-takeoff stories.
- Post-training is scaling toward pre-training's share of compute. Launay says recent models, Grok 4 among them, spent nearly as much on post-training as on pre-training. The host credits that focus with Grok 4's top Humanity's Last Exam score at the time of recording. Launay expects most modern models to get more post-training than pre-training, because its signals scale. Human thumbs-up feedback (RLHF) is limited, but verifiable or execution rewards, such as solving maths problems or passing tests mined from GitHub repositories, can be generated at massive scale. AI feedback from judge models covers tasks that can't be verified, such as checking a summary is factual.
- AI judges are about as reliable as human annotators. Launay says human annotation is noisier than people assume: agreement between two annotators is "crazy" low on many tasks, and model-to-human agreement lines up with it. He adds that verification is easier than generation, which he sees as a foundation of why AI feedback works, with the caveat that the judge must be capable at the task. The host reports that in his own experiment about a year earlier, GPT-4's ratings on a specialized task were indistinguishable from those of the human experts.
- RL generalizes better because it learns from the model's own attempts. Launay compares supervised fine-tuning to memorizing a worked solution in general relativity, which fails once the black hole carries a charge. RL has the student attempt the problem and get corrected. Because RL samples come from the model itself, they stay in its distribution and gradually reshape it, while SFT "plops down" a new distribution that may not connect back to existing knowledge. He says PPO and GRPO differ mainly in how they assign credit for a reward: PPO trains a value model, GRPO averages over several rollouts.
- RL's difficulty is engineering, which Adaptive packages as "RLOps." Launay says RL mixes inference with training and needs judge models (Adaptive runs pipelines with five to 10) and environments such as SQL databases, compilers or VMs. Pre-training, by contrast, scales easily because it is simple. Adaptive's tooling turns a Python recipe into a distributed run on open models such as Llama or Qwen. He claims RL always gives the best trade-off between cost and performance. In one unnamed client's chatbot project, about 70 human critiques and rewrites became about 80,000 synthetic conversations of roughly 10 turns each, which he calls "nearly 10 million data points" once candidate answers are counted. He names AT&T as a client whose teams now run the tool on their own.
- Superintelligence will arrive spiky and messy. Launay's bet is that superintelligent systems arrive within about five years, or 10 if you are bearish. Mathematics would come first because it can advance "in a box" with formal verification, while biology and medicine stay bound by real-world experiments, which he names as the next bottleneck to scale. He expects the transition to be turbulent, followed by "probably the best time ever." He also doubts photonic or quantum hardware is needed to get there: current silicon will build the AI that later helps build those.
Crosstalk SolutionsAre you Confused by UniFi OS Server? Let's Clear it Up!
- UniFi OS Server is for networks without UniFi OS on the gateway. The host sees two use cases: UniFi APs and switches behind a third-party gateway such as pfSense or an EdgeRouter, and UniFi gateways that lack onboard UniFi OS. Think of it as a cloud key, he says, running locally or cloud-hosted for many sites; Dream Machines, the Dream Router, Cloud Gateways and the Dream Wall don't need it.
- The gateways without onboard OS are aimed at MSPs. The store's advanced-hosting gateways (Gateway Enterprise, Gateway Fiber, Gateway Lite) suit a provider running one large cloud or bare-metal UniFi OS Server instance, giving one interface for all customer sites and one set of admin credentials per employee instead of per-site logins.
- It replaces the legacy self-hosted Network application and cloud keys. The host has moved about a dozen clients off an old Network instance hosted at DigitalOcean. Gains over the legacy app: UniFi OS and its applications update from the UI and can auto-update rather than via the CLI, and sites get UniFi Identity, InnerSpace and Site Magic SD-WAN mesh VPN.
- It runs only Network and cannot act as a router. As of recording it cannot run Protect, Access or Talk, only Network plus Identity and InnerSpace, though Ubiquiti has hinted at more. It does not turn a multi-NIC box into a firewall like pfSense. He assumes, untested, it can't share a server with the legacy app because of port overlap (the 8080 inform port, a new 11443 GUI port), so shut the old app down first.
- It is early access, and a Docker version is unlikely. Downloads appear only to those signed up for UniFi early access, which he thinks is free. He guesses Ubiquiti won't build a Docker version because the server is already containerized, in Podman.
Stock Tok$Quick Monday Morning Stock Update
- This post includes paid content that is not available publicly. See the original source.
Polis Pandit$A New Chapter For PolisPandit - AND an Invitation
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnWhat is Convex Optimization? (with Akshay Agrawal)
- Convex problems are the optimization problems solvable efficiently, reliably and provably. Agrawal defines mathematical optimization as assigning values to variables, such as how much to invest in each stock, to minimize cost or maximize reward under constraints like a budget or no short-selling. Convex optimization is the subset that can be solved "super efficiently, super reliably, provably."
- Convex optimization runs real portfolios, energy systems and rocket landings. Agrawal, a contributor to CVXPY, says many billions of dollars are allocated daily through portfolio optimization solved this way, and that it is used in energy management and real-time control. He says SpaceX uses software developed by his lab to land rockets ("I think it's like the Falcon or something"). These are cases where the world can be modeled with understandable constraints whose structure can be exploited.
- Machine learning overlaps with convex optimization but has different goals. Agrawal notes that classical models such as logistic regression and SVMs are fit with convex optimization under the hood. ML usually has few or implicit constraints, offers less insight into what the model does, and seeks a good-enough solution that generalizes to unseen data rather than a provable optimum. His embeddings library, PyMDE, which fits embeddings on a GPU, uses convex optimization techniques to solve an ML problem efficiently.
Dr. Jon KrohnThe Fatal Flaw in Jupyter Notebooks (and How Marimo Fixes It)
- Jupyter's hidden state silently breaks reproducibility; marimo catches it immediately. The host cites a study, which Agrawal has referenced before, finding that only 4% of Jupyter notebooks reproduce their original results. Agrawal says Jupyter suits quick REPL-style hacking but is relied on where reproducibility is paramount. Because it doesn't re-run dependent cells, it leaves the user to track dependencies. If you delete the cell defining a PyTorch model class, the class stays in memory and the notebook keeps working until a fresh run the next day fails. In marimo, deleting that cell removes the variable from memory and invalidates the cells that use it.
- Notebooks can carry their own package environment. Agrawal describes marimo's opt-in "sandbox" mode, which records each installed package and its version in the notebook file itself. Re-running the notebook creates an isolated virtual environment with just those packages, so a single file can be sent to someone and run without setup.
Technology and FriendsJennifer Marsman on NLWeb [EPISODE 864]
- NLWeb aims to put LLM-style questioning on every website. Marsman, a principal engineer in Microsoft's office of the CTO, describes it as both an open web protocol and an open-source implementation, recently moved from Microsoft's GitHub to its own organization because it "doesn't need to be run by Microsoft". Its founder, Guha, who invented RSS and RDF, frames it as part of an "agentic web" usable by humans and agents alike; her own complaint is a grocery site that returns refried beans for a green-beans search.
- It answers by retrieving from your own catalog, curbing invented items. Data comes in from any RSS feed or a specific JSON format, gets embedded into a vector database, and queries run retrieval-augmented generation over it; it works best on collections such as recipes, playlists or product inventories, leaning on schema.org-style structure. She says it will always return something from the actual catalog, while conceding it is not 100% hallucination-free.
- Embeddings, model and vector database are all swappable. Several providers for each are supported, including Qdrant and Azure AI Search for storage, and none needs Microsoft technology. She believes Qdrant is the default store because it runs locally, is unsure of the other defaults, and notes that someone from Hugging Face contributed support for any Hugging Face model.
- Every NLWeb site automatically exposes an MCP server. The server, called "ask", takes natural language and answers from the site's data, so any MCP client, such as Claude Desktop or VS Code, can query it. Guha's comparison is that NLWeb is to MCP as HTML is to HTTP: separate layers that should evolve independently.
- Partners are trialling it, but she stops short of calling that production. It was announced in Kevin Scott's keynote at Build 2025 with travel, recipe and retail partners; when Giard said companies were using it in production, she answered "I wouldn't say production production". She invites outside contributions and says the team is hiring a senior and a principal engineer.
Dr. Jon KrohnNotebook to Full Web App in Seconds with Marimo
- A marimo notebook becomes a shareable data app without migration. Agrawal says data scientists previously needed a front-end engineer, or hit performance bottlenecks with tools like Streamlit, to make an experiment actionable. In marimo, they swap a few variables for sliders, dropdowns or tables and share the result with their team or CEO. He says CEOs at some companies use notebooks their data scientists built, and some write their own, helped by the LLM integration.
- Storing notebooks as Python files eases the handoff to engineering. Every marimo notebook is an executable script that runs with
python my_notebook.py, accepts command-line arguments, and can run as a pipeline or cron job. Other code can import functions and classes from it, and it works with pytest. Agrawal argues this streamlines the long-discussed handoff from researcher to engineer, because marimo nudges users toward reusable functions.
GeoHussarThe closest that the world ever came to nuclear war was in 1983, during NATO operation able archer
- Able Archer was the world's closest brush with nuclear war. He places it in 1983 rather than the 1960s and says nobody knew at the time. NATO's annual exercise involved tens if not hundreds of thousands of troops, and the 1983 edition went all out: every branch, tanks, jets, simulated nuclear weapons, real-style communications and world leaders reading scripts.
- The Soviets read the simulated launch as cover for a first strike. Seeing erratic high-level communications and large formations, the Soviet military raised readiness for the first 24 to 48 hours and moved fighters nearer the border. The day of the simulated launch caused panic, and he says the Soviets were right to be alarmed; the captions cut off there.
Technology and FriendsSteve Smith on How Generative AI is the New Offshoring [EPISODE 865]
- Replacing developers with AI repeats the early-2000s offshoring mistake. Smith says companies offshored to developers costing about one-tenth of US rates, then reshored after quality fell, owing to time zones, communication and culture rather than skill. Fast copy-paste code built technical debt until a feature that took a week in year one took three months by around year five, plus three more of bug-chasing.
- The hidden cost is maintainability, not rework. Giard recalls clients arguing they could redo cheap work nine times and still profit; Smith answers that unmaintainable code makes every next feature cost more, growing "exponentially", and that rewriting under live customers is hard whether agents or humans do it.
- Agents excel at small greenfield work but falter on enterprise complexity. He says a to-do app that would take him 10 hours takes one with coding agents, but on a large codebase he can spend eight hours going back and forth, and agents sometimes do exactly what the prompt forbids. Unlike a junior developer, the AI does not learn from that time, and friends running agents on five monitors become the bottleneck reviewing output that a non-programmer could not judge.
- Today's $20-a-month pricing is unlikely to last. His forecast: vendors are spending investor money to win market share and cannot profit at that price, so costs could reach $200 a month (which he says Claude already charges), $2,000 or $20,000, at which point companies will ask how many developers the bill would have paid for.
- Use AI for grunt work and experiments, not as a panacea. He calls it promising and uses it constantly, for example to summarise half a dozen changed files into a commit message. His article on the subject uses Fantasia's sorcerer's apprentice, drowning in water his brooms fetch, as the image of teams drowning in poor AI-written code.
Dr. Jon Krohn912: In Case You Missed It in July 2025 — with Jon Krohn (@JonKrohnLearns)
synopsis
The host's July compilation strings together five guest clips around one theme: how to know whether a model, its data or its predictions can be trusted. It ends on causal modeling as a way past correlation.
Data first. A legal-tech researcher says noisy labels made her algorithm comparisons meaningless, which led her to data-centric ML: hold the model fixed and engineer the data. That work runs through MLCommons' DataPerf, now a DMLR working group, journal and workshop series, with a Common Crawl challenge to build a low-resource-language dataset.
Judging models. A benchmarking guest says benchmark answers always leak, and that Chatbot Arena's blind votes capture preference rather than knowledge, and notes the alleged arena-tuned Llama 4 variant. A Bloomberg researcher adds that guard models such as Llama Guard target broad consumer risks, not regulated finance, and recommends domain-adapted risk taxonomies such as NIST's, red-teaming events and evaluation in the deployment context.
Predicting people. A predictive-analytics founder cites decades of replicated neuroscience showing that decisions start before awareness, with fMRI predicting turns 10–15 minutes ahead in a driving simulator. The founder argues that customer behavior is just as predictable from historical data.
Causal tools. Robert Ness closes by arguing that libraries such as Pyro, PyTorch and DoWhy handle the inference, so modelers can concentrate on stating causal assumptions.
- Data-centric ML fixes the model and engineers the data instead. A legal-tech guest says label noise so badly undermined her algorithm evaluations that she moved into data-centric AI, the term she credits to Andrew Ng. She joined MLCommons' DataPerf working group, which led to workshops at ICLR and ICML, a NeurIPS paper, a DMLR journal, and a DMLR working group. That group is partnering with Common Crawl on a challenge to produce a public low-resource-language dataset. Her examples of the approach include DoReMi, which weights domains of the Pile for better LLM pre-training, and selecting which data points to label before any training.
- Blind arena voting measures preference, not knowledge. A benchmarking guest argues that benchmark answers inevitably end up online, and that hiding them raises the question of who owns them. The host suggests a Chatbot Arena–style setup, but the guest says arena votes are essentially thumbs up or down, which OpenAI blamed for sycophancy. The guest also relays, stressing it is only an allegation, that Meta tested an arena-specific Llama 4 variant that differed from the released model.
- Off-the-shelf safeguards miss domain-specific risks, so evaluate in context. A Bloomberg researcher's paper uses capital markets and asset management as a case study. Its point is that foundation models and guard models such as Llama Guard and ShieldGemma are trained for broad consumer risks, not for regulated, knowledge-intensive domains such as finance, healthcare or law. The researcher's recommendations: adapt general taxonomies such as NIST's AI risk management framework, run red-teaming events to measure how often a system gives financial advice or refuses, and don't trust providers' benchmark scores to carry over.
- Unconscious brain processes make human behavior predictable in advance. A guest from episode 907 calls it agreed fact, replicated since experiments in the 1980s associated with Benjamin Libet, that brain processes we can't consciously access drive our decisions. The guest says fMRI can predict a driving-simulator turn 10 to 15 minutes ahead. They argue businesses can likewise predict purchases, churn and lifetime value from historical behavior, and say their company's mission is to bring that to small and midsize businesses.
- Causal libraries free modelers to focus on assumptions. Robert Ness, guest of episode 909, says his book separates causal reasoning from statistical and computational machinery. Just as Stan handles Hamiltonian Monte Carlo for continuous models, tools like Pyro, PyTorch with a differentiable loss, and DoWhy handle the inference. Modelers can then toggle between estimators such as regression, propensity scores or double ML to see which gives stable results.
Scott HanselmanLearning to Code with AI and Steve Klabnik
synopsis
Klabnik, a software engineering manager at Oxide and co-author of the Rust book, ends up as a willing user of AI coding agents who keeps the professional duty of review. He says he is "definitely having fun": Claude now writes most of his code, he works at the level of architecture and design, and he reviews whatever he ships. The cost, he says, is social rather than technical, since friends who oppose the tools have turned on him. Hanselman, who calls himself "a bit of an AI vegan" who chooses what AI he accepts, largely agrees, while pressing on what is lost.
Where each stands. Klabnik says he was firmly anti-AI after early versions looked like a parlour trick, until Claude 4 prompted him to try again "to be an informed hater". He finds coding a different question from other uses.
Harms. On Grok's "spicy mode" producing images of Taylor Swift, Hanselman asks how such output got past filters that, at Microsoft, sit in front of everything. Klabnik's larger worry is deepfakes of ordinary people at scale, such as scam calls to his 80-year-old mother. He compares the moment to the fear that the web would end literacy: society partly adapted, he says, and he is cautiously optimistic about human resilience.
Copyright and credit. Asked whether training on an artist's 20 years of style is fair, Klabnik recounts the commission that followed the 1978 copyright act and decided source code was more like a book than a recipe. He argues that stronger copyright would stifle the way music scenes build on each other, and that licences miss the social expectations of open source. Hanselman adds his years as an expert witness, with the EFF, against a patent troll who sued tens of thousands of small sites.
Working with Claude. Klabnik describes himself as "an open-source project of one". He no longer enjoys a whole Saturday spent debugging and cares instead about IR choices, optimization passes and semantics. He uses Claude Code because he wants agentic tools, not Cursor's autocomplete focus. His compiler, written by hand without LLVM, currently supports only integers, booleans, empty tuples, if, while and function calls, and he is adding arrays and structs. He insists a one-shot prompt cannot build a compiler.
Smart or dice roll. His two stories make the point. In one, Claude ran GDB and objdump unprompted and fixed a segfault. In the other, it computed struct offsets from field names and then tried to return to that hack. Hanselman calls the tool a pachinko machine. Klabnik, around a two-year-old who repeats everything she hears, is unsure what "thinking" means.
Side effects. Klabnik reports a paper finding that the tools fix bugs worse when variable names are gibberish, and Richard Feldman's case where Zed updated a code comment and so exposed a behaviour the refactor broke. He concludes that LLM documentation beats none, and that what helps LLMs often helps humans.
- A former AI skeptic now has Claude write most of his code. Klabnik, a software engineering manager at Oxide and co-author of the Rust book, says he was "very anti" until the release of Claude 4, when he retried it to be "an informed hater" and found it good. He says Claude has written most of his code for about two months; he reviews every PR he merges on professional work as a matter of responsibility, though not on some low-stakes personal projects. He prefers agentic Claude Code to Cursor, which he sees as built around autocomplete.
- AI moves his interest up a layer, to architecture. He likens it to auto-formatters: he now cares about IR design, optimization passes, language semantics and tests, not whether a function is 20 lines or two 10-line ones. A hand-built compiler without LLVM, a project he says he would not have started without these tools, is being written almost entirely by Claude. Hanselman's parallel is a podcast backend he had put off for years and wrote in 42 minutes, which leads him to ask whether he has cheapened the value of software.
- The agent alternates between impressive and absurd. Klabnik says that on a segfault Claude independently ran GDB and objdump, found the bug he and a friend had found plus another, and fixed it. Yet it also calculated struct field offsets by pattern-matching field names, and after agreeing to fix that properly it reverted to estimating from names. Hanselman's analogy is a pachinko machine: "a miracle", but "not smart"; Klabnik is less sure what "smart" means.
- Deepfakes of ordinary people worry him more than celebrity cases. On Grok's "spicy mode" producing nude images of Taylor Swift, Hanselman says at Microsoft nothing ships without a responsible-AI filter in front. Klabnik calls the Swift case bad but says deepfakes at scale of people without public lives, such as fake kidnapping calls to his 80-year-old mother, are the larger social problem, while voicing cautious optimism about human resilience.
- Copyright is an awkward tool for AI-era creativity. Klabnik recounts that a commission after the 1978 copyright act weighed whether source code was like an uncopyrightable recipe or a book, and sided with the book. He argues stronger copyright would hurt the way scenes form (requiring the first punk band to be paid would have prevented punk), and that licences are not the whole story of open source. He says his refusal to disavow AI tools has cost him friendships, and that LLM-written docs, while worse than finely crafted ones, beat none.
GeoHussarWas it #Christianity that made #Europe so violent? #history
- Pre-Christian Rome already fought a major war about every decade. Answering a commenter, he counts at least 50, some say 60, wars from the Republic's start around 500 BC to the first century, roughly one every 10 years over 500 years.
- Christian Byzantium kept the same pace, so Christianity isn't the cause. Rome became officially Christian in 380 AD under Theodosius; setting aside the Western Empire, which fell in 476, he counts about 100 wars for the Eastern Empire from 395 to 1453. A war every decade before and after leads him to conclude Christianity did not make Europe violent.
Crosstalk SolutionsUniFi OS Server on Raspberry Pi: Here's Your Easy Setup Guide!
- A Raspberry Pi suits UniFi OS Server for your own network only. The host frames this install as controlling local devices without exposure through the firewall; MSPs hosting customers should use his hardened hosted-Linux setup instead. It serves UniFi APs and switches behind a non-UniFi gateway, or UXG gateways lacking onboard UniFi OS; Dream Machines, the Dream Router, Dream Wall and UCG models don't need it.
- The hardware minimum is a guess; storage is the firm one. Ubiquiti publishes no CPU or RAM minimum, and he guesses a Pi 4 with 4 GB will do; he used a Pi 5 with 8 GB. Storage must be at least 20 GB, so he recommends a card of 32 GB or more, and used 128 GB.
- Ubuntu Server plus Podman and the ARM64 installer is the path. He flashed Ubuntu Server 25.04 with Raspberry Pi Imager, enabling SSH with password authentication up front, and advises a DHCP reservation or static IP since devices check in regularly. After updating, install Podman,
wgetthe Linux ARM64 installer,chmod +xit, run it withinstallunder sudo, and allow about two minutes before browsing to port 11443. - Ubiquiti's licence forbids reselling hosting built on the free software. The setup wizard states it is for personal or organizational use only. The installer's note also says to add your user to the
uos-servergroup to run its commands; a factory-reset U7 Pro XG then appeared ready to adopt. - MicroSD cards are prone to failure, so plan for it. He notes UniFi backs the configuration up to the cloud weekly, so a failed Pi can be rebuilt and restored; alternatively boot from an external M.2 or SATA SSD, copying the boot volume over before installing UniFi OS Server, which he says is easiest.
Technology and FriendsGuy Royse on Ham Radios [EPISODE 863]
synopsis
Royse presents ham radio as a century-old, decentralised, still-technical hobby whose real point is resilience: with a licence, a battery, a radio and a wire in a tree, an operator can talk around the world and carry messages out of a disaster when the internet and cell networks fail.
Where the name comes from. "Ham" derives from telegraphy: an operator's keying style was their "fist", and poor amateurs were "ham-fisted". Royse uses the ARRL, the hobbyists' advocacy body, and its annual handbook to show how old the culture is.
What operators do. Beyond listening and casual "rag chewing", his club runs a weekly net in which a net controller logs call signs and calls on each operator in turn, since only one can talk at a time. Traffic nets relay radiograms verbatim, which he frames as emergency practice; Giard raises the 9/11 internet overload, and Royse says emergency services also reserve part of the cell network.
How signals travel. Antennas run about half a wavelength. VHF and UHF reach the horizon; HF bounces off the ionosphere, so a 40-metre antenna is 20 m of wire strung into a tree. Reach depends on conditions.
Hardware. A $25, 5-watt Baofeng reaches the horizon, or 60–70 miles via Columbus's main repeater; his roughly $1,200, 100-watt rig adds HF. The radios are analog in the sense a CD is, with digital modes layered on top, and Morse code's narrow bandwidth lets about 20 watts reach the world against about 100 for voice.
Getting in. He came through web SDRs and licensed to join the data modes. He recommends YouTube, chosen for context, and ARRL.org for clubs and exams.
- "Ham" comes from telegraphers' "ham-fisted", not an acronym. Royse, a developer advocate at Redis, explains that telegraph operators called an operator's keying style their "fist", so unskilled amateurs were "hams", a slur the community later adopted. The hobby's age shows in terms like "rag chewing" and "Elmer" (a mentor), and in the ARRL handbook, published yearly for over a century; his copy is the 100th-anniversary edition.
- Message-relay nets are practice for disasters. In traffic handling, operators relay a "radiogram" word for word through several intermediaries to its recipient, which Royse says lets someone in a hurricane zone get "I'm alive" messages out with only a battery and a radio. Giard raises websites failing on 9/11; Royse adds that emergency services also lock down part of the cell network in a crisis, shrinking capacity just as traffic surges.
- A battery, a radio and wire in a tree can reach worldwide. Antennas should be about half a wavelength: 2-metre signals (144–148 MHz in the US) reach only the horizon, while lower HF frequencies bounce off the ionosphere, so a 40-metre-band antenna is 20 m of wire, hoisted with an arborist's throw or his club's air-powered tennis-ball launcher. He says a 100 amp-hour battery lets him transmit for two days, and he reaches the eastern US easily from central Ohio, depending on ionospheric conditions.
- Gear ranges from a $25 handheld to a $1,200 HF rig. His 5-watt Baofeng handheld covers 2 m and 70 cm and reaches the horizon, or about 60–70 miles through Columbus's biggest repeater atop a state office building; repeaters are usually bought by local hams and placed on towers such as a fire department's. His 100-watt rig adds HF, and he says Morse code needs less bandwidth than voice's roughly 3 kHz, so about 20 watts reaches anywhere on CW against about 100 for voice.
- Start with YouTube and the ARRL. Royse came in through web-hosted software-defined radios, then needed a licence to join the digital modes he heard; he likens radio to vinyl, a decentralised, physical way of connecting. He warns that many videos are "all trees, no forest", and points to ARRL.org for clubs, licence exams and practice tests.
Dr. Jon KrohnHow Reactive Notebooks Fix Jupyter's Biggest Flaw (with Marimo's Akshay Agrawal)
- Reactivity keeps code and outputs in sync, like a spreadsheet. Agrawal's example: with cells
a = 1,b = a + 1andc = a + b, changingato 2 and running that cell automatically re-runs the other two. He says this speeds exploration, and more subtly protects reproducibility, since a forgotten dependent cell would otherwise make every output untrustworthy. The host cites 4 million downloads at the time of recording, and a user called the nudging "gentle parenting." - Lazy mode guards expensive cells without losing guarantees. For cells that kick off costly GPU training or API calls, Agrawal says marimo can be set to "lazy" execution. Dependent cells are then marked stale rather than run automatically, which keeps the state guarantees without accidentally spending credits.
- Reactivity makes interactive widgets actually useful. Agrawal says Jupyter sliders need callbacks to pass values back to Python, and dependent cells still don't re-run. In marimo, replacing a variable with a slider re-runs downstream cells as it moves. An example he wanted from his PhD: selecting a cluster in a scatter plot with the mouse sends those points back to Python as a dataframe.
Master InvestorThomas Peterffy: What he learned from growing up in post-war Hungary
- Peterffy says censorship in socialist Hungary shaped his view of capitalism. Barred from current information and limited to classic authors like Dickens, Balzac, and Zola, he says those books taught him that a successful business succeeds by meeting people's needs better and more cheaply than competitors.
Stock Tok$Buying a Stock - Quick Update 🌩️
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnWhy Jupyter Notebooks Fail at Reproducibility (with Akshay Agrawal)
- Jupyter's JSON format blocks code reuse and clean version control. Agrawal, who did a Stanford PhD on vector embeddings and dimensionality reduction from 2018 to 2021, credits Jupyter for making that visual, iterative research feasible. But because notebooks are JSON blobs, code can't easily be imported across them, so researchers duplicate notebooks for each variant experiment. Running a notebook without changing code can still produce a large git diff from changed images or metadata.
- Out-of-order execution leaves notebooks in a hidden, untrustworthy state. Agrawal says that in science or data engineering, where determinism matters, running cells in any order is a flaw rather than a feature. Forgetting to re-run a dependent cell puts memory in "some like strange sort of quantum state," and you may learn "four hours later, eight hours later" that no results are valid. The root cause is that the runtime doesn't know how cells relate, unlike Excel, which recalculates.
- Untracked package versions stop others from reproducing results. Agrawal says researchers often forget to record exact versions, for example of PyTorch and transformers, in a requirements.txt or pyproject.toml. Colleagues who receive the notebook then can't reproduce the results saved in it.
Crosstalk SolutionsUnifiOS Server Migration | Export, Import, Done
- Migrating legacy UniFi sites to UniFi OS Server is easy, the host finds. His old DigitalOcean-hosted controller was still on Network 7.4.162, over a year since its last update, because the legacy app needs manual CLI updates; he is moving his legacy clients so they stay current. The use cases he gives are UXG-line gateways without onboard UniFi OS and UniFi devices behind a non-UniFi gateway.
- Export the site, import it, then push devices across. On the old controller, Settings > System > Site Management > Export Site downloads a file; on the new server, add a site via Import Site. Back on the old controller, pick the devices and enter the new server's hostname (the default port needs no change), and they are told to set-inform to it.
- Devices moved over quickly once told the new address. Within one to two minutes, an old switch and a UAP AC Pro reported to the new server almost immediately, and he believes they are also updated to firmware matching the new server's version. Only then forget the devices on the old controller and delete the site.
- It is still early access, so production MSPs may wait. As of recording, he suggests those with many production clients hold off until general availability. Older hardware can get stuck; then SSH in and run
set-informmanually, or paperclip-reset the device and re-adopt it.
Master InvestorThomas Peterffy: Why people misunderstood Trump's tarriffs
- Peterffy says he saw Trump's tariff logic as rational, not destructive. He says institutional investors misread Trump in an "echo chamber" of panic, while he argued publicly on April 6 — which turned out to be the market's low point — that the sell-off was the greatest buying opportunity of his lifetime.
- He says he isn't worried about a pullback despite the rally. He expects "smooth sailing" for the market for some time to come.
Dr. Jon Krohn911: The Future of Python Notebooks is Here — with Marimo’s Dr. Akshay Agrawal
synopsis
Akshay Agrawal, co-founder and CEO of marimo, argues that a Python notebook can be reproducible by construction and still be as interactive as Jupyter. His design: a reactive notebook stored as a plain Python file that doubles as an app, a script and an importable module. That shortens the path from exploration to deployment and lets data scientists ship work without a front-end engineer. The open-source tool, at 4 million downloads by the time of recording and backed by AIX Ventures, is free for individuals. The business he plans is enterprise-scale experimentation on very large datasets.
Origins. Agrawal's 2018–2021 Stanford PhD on vector embeddings depended on Jupyter, but he hit its limits. JSON files blocked code reuse and made git diffs noisy. Out-of-order execution left hidden state, with results that might turn out invalid "four hours later, eight hours later." Untracked package versions broke handoffs to colleagues. After the PhD, in 2022, he surveyed the field and found Julia's Pluto notebook, which already had reactive execution, UI elements and plain-file storage. He wrote a 2,500-word design document and refined it with researchers at Stanford, SLAC and a few startups before raising money.
Reactivity and reproducibility. Change a and the cells using it re-run, like a spreadsheet. Agrawal says this speeds exploration and guarantees that outputs match the code. A lazy mode marks dependent cells stale rather than running them, to protect costly GPU jobs. Deleting a cell removes its variables from memory, so errors surface right away. The host cites a study, previously referenced by Agrawal, finding that only 4% of Jupyter notebooks reproduce. Agrawal concedes Jupyter fits quick REPL-style hacking. An opt-in sandbox records package versions in the file and builds an isolated environment on each run.
Notebook, app, script. Sliders and selectable plots feed values back to Python. Hiding the code yields a data app that re-runs only dependent cells (he contrasts Streamlit's performance limits) and can run in the browser via WebAssembly. Because notebooks are Python files, they run from the command line, schedule as cron jobs, are importable and testable with pytest, which Agrawal says eases the handoff from researcher to engineer. He reports CEOs using, and sometimes building, marimo apps. The AI assistant injects an in-memory dataframe's schema and samples into prompts, something he says Cursor can't do.
Product and community. Early features came from the founders' gut. Now the test is "big rocks," features that boost productivity a lot for many kinds of user. On community, he credits kindness, openness and same-day bug fixes, which he says win "a supporter for life," citing Astral's uv as a model. First-time contributors can start with documentation fixes or issues tagged "good first issue."
Optimization roots. Half his published research is on convex optimization, the subset of optimization problems solvable efficiently and provably. He says it allocates many billions of dollars daily through portfolio construction, runs energy management, and helps SpaceX land rockets with software from his lab. He contrasts this with machine learning's search for good-enough solutions under implicit constraints. His aim, he says, is making mathematics "actionable."
- Jupyter's design undermines reproducibility in ways marimo targets. Agrawal, marimo's co-founder and CEO, drew on his 2018–2021 Stanford PhD. Notebooks stored as JSON blobs block code reuse and produce noisy git diffs. Running cells out of order leaves "hidden state," for instance a deleted cell's model class lingering in memory until a fresh run fails. And untracked package versions stop colleagues from reproducing results. The host cites a study Agrawal has referenced finding only 4% of Jupyter notebooks reproduce their results. Agrawal says Jupyter still suits quick REPL-style hacking.
- Reactive execution keeps outputs in sync and makes notebooks into apps. Agrawal says that, as in a spreadsheet, changing a variable re-runs the cells that depend on it. A "lazy" mode instead marks those cells stale, for expensive GPU or API work. Sliders and plots feed values straight back to Python, so selecting points in a scatter plot returns a dataframe. Hiding the code turns a notebook into a data app. He says it performs better than Streamlit because moving a slider re-runs only what depends on it, and it can be served or run in the browser via WebAssembly.
- Notebooks stored as Python files ease the researcher-to-engineer handoff. Each marimo notebook runs as a script with command-line arguments, works as a pipeline or cron job, can be imported as a module, and works with pytest. Agrawal says some CEOs use notebooks their data scientists built, and some build their own. Its sandbox mode records package versions in the file and rebuilds an isolated environment on each run.
- Its AI assistant sees live data that a code editor can't. Agrawal says marimo's generate-with-AI feature, which writes Python or SQL using OpenAI, Anthropic, Gemini or local models through Ollama, can be pointed at an in-memory dataframe. It then pulls the dataframe's schema and sample values into the prompt so generated code uses the real column names. He argues tools like Cursor, working on text files, can't do that.
- Open source stays free; revenue will come from enterprise scale. Agrawal says open-source marimo aims to be the best environment for individuals working with their own data. He plans to commercialize the unsolved problems of rapid experimentation on very large enterprise datasets. The host likens this product-led growth to RStudio's path. AIX Ventures, whose investor Sean Johnson introduced the guest to the host, has invested in the company, and the host cites 4 million downloads at the time of recording.
Mr. Global🔥 A Global Energy Conflict Is Brewing 🔥 PLEASE SHARE!
- OPEC's output hikes are meant to bankrupt Russia through cheap oil. OPEC agreed a further 547,000 barrels a day increase for September, bringing this year's increases to over 2.2 million barrels a day and restoring all prior cuts. He reads the Trump administration's pressure for those increases as an effort to crush the oil revenue funding Russia's war in Ukraine.
- Prices failed to fall because China is soaking up the extra supply. He says China bought over a million barrels a day of additional oil through the first half of 2025, storing rather than using it. Its strategic petroleum reserve is essentially full, but he puts storage outside it at 67% of capacity, enough to keep buying for at least another 6 months. A weak dollar helps prices, he grants, but cannot explain the gap.
- China's buying looks like cooperation with Russia, not bargain-hunting. His inference: a buyer only after cheap oil would have let the added supply crash the market and bought later at a deeper discount. He frames China, India, Russia and Iran as working against US efforts to end the war.
- The China and India tariff threats are about energy, not trade. Trump has threatened China with sanctions over its purchases; India increased Russian oil buying after tariff threats and refused to stop, and he expects significant tariffs on India "any day, possibly today." He is unsure tariffs will work.
- He expects the glut eventually to overwhelm buyers and sink prices. He forecasts a significant price drop, timing unknown; many expect early 2026, and he expects winter demand to fall. He notes OPEC timed its increases to seasonal demand peaks, regaining market share while doing Trump a favor.
Polis Pandit$There's a Slim Shady in all of us
- This post includes paid content that is not available publicly. See the original source.
Dr. Jon KrohnHow to Build a Causal AI Model, with Robert Osazuwa Ness
- A causal workflow starts from a question and explicit assumptions. Ness's example question is whether side-quest engagement drives in-game purchases. He encodes his assumptions as a DAG, for instance that guild membership causes both engagement and purchases. He specifies the DAG in code, for example as a NetworkX object, and passes it to a library such as DoWhy to get a model object.
- Identification asks whether your data can answer the question. Given the DAG and the variables actually observed, possibly with unobserved confounders, identification tests whether the effect can be estimated at all. If not, Ness says you look for additional variables. One option is an instrument, which affects purchases only through engagement and enables instrumental-variable analysis. Another is a mediator between cause and outcome, which enables front-door analysis.
- After identification, estimation is ordinary statistics, followed by refutation. Ness says the "causal heavy lifting" is done by then. Choosing between linear regression, propensity scores or double machine learning involves the usual trade-offs, for example wide confidence intervals in one method and instability in another. He ends with sensitivity analysis, which DoWhy calls refutation, to test how robust the conclusions are to missed variables, small data or leaning on linearity.
Mr. GlobalIt’s About Water #truth #breakingnews #news
- Clean water, not medicine, did most for life expectancy. After a Neil deGrasse Tyson clip saying half of early humans died before 30, he grants that medicine matters but says clean water has always been the biggest contributor. He then points to Flint, Michigan and implies, without stating it, that places lacking clean water share the race of the people who live there.