Neward argues that precise, plain prose, written before any presentation, is a core engineering skill that school never teaches and seniority demands. His prescription: write the argument out, strip judgment words and repetition, use the active voice, and ground claims in data.
Why it matters. Neward, a liberal-arts graduate now at Capital One and a guest lecturer at the University of Washington, says the higher an engineer goes, the more the job is influence through writing. Giard frames poor communication as a ceiling on a code-only career.
The six-pager. Neward's talk opened with silent reading, borrowed from Amazon's six-pager: six pages fill 20 minutes of reading, the first third of a meeting. He holds that talking through an idea without writing it is intellectually cheap.
Precision. Delete adjectives and adverbs, restoring one only to resolve ambiguity; "simple" and "easy" are value judgments that alienate readers who disagree. Say things once: unlike a speaker, a writer gets no live feedback, and readers can reread.
Voice and tone. The passive voice ("Mistakes Were Made") obscures how a failure happened; blameless culture can use the active voice. Avoid sarcasm and prose that draws attention to itself.
Data and anecdata. Neward's Maven example shows a few complaints prompting a measurement, and he cites a study of GitHub commit logs that found C draws the most profanity.
Kirill Eremenko, who founded both this podcast and the SuperDataScience.com platform, walks through the eight weeks of the platform's first AI engineering bootcamp cohort as a recipe listeners can follow on their own. His position is that an AI engineer stands out by understanding the whole path end to end: building a proof of concept that solves a defined business problem (weeks 1–4), then deploying it so it is secure, reliable, scalable and cost-effective (weeks 5–8). The course ends with a self-improving "digital twin" agent running on AWS with CI/CD. He says the deployment half exists because job candidates kept being asked about deployment in interviews, which other bootcamps didn't cover.
Prerequisites and format. Participants needed Python, some PyTorch, scikit-learn and pandas, LLM API calls and basic AWS experience such as launching an EC2 instance from the console. Cohorts are capped at 10 people. Each week has a three-hour interactive core session, two hours of office hours on commercial use cases, and expert AMAs. Ed Donner taught the first half, on what is possible. Sam, an AWS specialist, taught the second, on what deployment allows.
Weeks 1–2: mindset and behavior design. Week one counters the hype: many problems need a script or robotic process automation rather than an agent. The tip is not to build anything until the business goal and success metric are defined. Participants compared 13 LLMs. When Eremenko asks about chat versus reasoning models, the host explains the difference through Kahneman's fast and slow thinking: reasoning models check their steps, spending more compute to cut errors. Week two treats prompt engineering as the designer's job: templates, parsing JSON output, and a Gradio front end for a flight assistant. Eremenko's tip is that the system prompt is always sent, even when the conversation overflows the context window. The host adds that models can't attend evenly across a million tokens, so needle-in-a-haystack results overstate recall.
Weeks 3–4: RAG and agents. Eremenko argues the field has moved from fine-tuning to customizing at inference time with RAG, because fine-tuning is costly, models change often and closed models can't be tuned. Participants built a pipeline with OpenAI embeddings, Chroma and LangChain, studied hierarchical retrieval, and learned to chunk with context in mind so an answer isn't split across chunks. Asked whether RAG survives contexts of 1 to 10 million tokens, both say yes: retrieval is more accurate, and a huge context is costly to resend every turn. In week four, Eremenko stresses that an agent's LLM only requests tool calls in text, and the developer's code should execute them and decide the routing. The project was a digital twin built from each participant's career materials. One participant used theirs live in a job interview and was hired.
Weeks 5–6: production and security. Week five moved the app into Docker on AWS, with infrastructure as code and CI/CD. It ran on Lambda rather than EC2 because occasional use tolerates a cold start. The tip is to put a caching layer in front of the LLM, which Eremenko thinks suits about 80% of business applications. Week six added long-term memory and a red-team exercise: teams got their own deployed agent to refund economy tickets against its rules, for instance by renaming a ticket as "special." The lesson is not to rely on prompts for security, but to enforce rules in code and log and verify tool calls before sensitive actions. He adds that MCP, Anthropic's popular tool standard, isn't yet secure enough for production because MCP servers have been hacked.
Weeks 7–8: production RAG and the capstone. Because embedding models are shallow, Eremenko recommends an LLM rewrite each query several ways and rank the results, so that "PTO" can find a policy written as "prorated annual leave." The capstone deploys the digital twin through GitHub Actions into Docker on Lambda. His principle is to treat it as a product, not a project: when the twin can't answer a question, it logs the gap and asks its owner to update the knowledge base. Asked to critique the curriculum, the host says the first half is as good as it could be. He calls himself less expert on the deployment half but finds it at least directionally correct.
pip install on generated code without checking it pulls the malware into their project.Sullivan, the solo developer at Honeybee Games, ends up attributing Focus Friend's rise to number one on both app stores to two things: a year of user research and careful iteration, and a deliberate refusal to monetise through ads, data collection or psychological tricks, which she argues is part of why people responded to it. She says she did not expect it: 100,000 users would have been a win.
Origin. After a January 2024 dinner where she said creators could sustain themselves with apps rather than merchandise, Hank Green contacted her on TikTok the next week. She proposed a focus timer; he proposed a magical bean. Her complicated pregnancy then delayed the launch by about a year.
Design and values. The gameplay loop rewards focus time: the bean knits socks, and the socks buy decorations, replacing an earlier build-the-room idea that ran out after about 13 items. She says she wants habit-building that helps people, not daily compulsion, so purchases are cosmetic. The art style came from interviewing about 40 focus-timer users, and the animation is skeletal rigging in Spine.
Launch and stability. A soft launch at the end of July via a YouTube community post produced a few thousand downloads and corrections to her assumptions, after a year with about 100 Discord beta testers. With no networking, no ads and few libraries, little can crash. The remaining crash is an iOS 15 Screen Time authorization difference, and an earlier bug locked users out of their phones.
The Instagram bug. Her hardest bug, which she could not reproduce, turned out to be Apple's "ignore for today" on app limits granting a day-long system pass. She also explains the always-allowed list for medical apps, which Hanselman checks for his diabetes apps, and the conflict with Opal.
Reputation over revenue. Citing the Panels backlash, she persuaded Green that ads, even opt-in ones, mean collecting data, so the app has none.
Michelle Yi, an AI entrepreneur, investor and researcher who started at IBM Watson at 16, argues that trustworthy AI depends on securing both the model and its data, and that attacks on both are now cheap and easy. Her position is that it is "not too late" to fix this. Whether enough money will go into it is "questionable." The fixes she backs are systematic red teaming and evaluation, defenses that work at the level of the model rather than case by case, and world models that can simulate consequences.
Red teaming and evaluation. Yi says both are routinely dropped because they slow deployment, and that this is why teams often see no ROI from a deployed proof of concept. Dedicated, diverse, technical red teamers find out-of-distribution cases, such as a rare illness getting a generic "take ibuprofen" answer. They build benchmark subsets and look for deliberate poisoning and jailbreaks, which she expects to matter more as vision-language models reach production. Constitutional defenses and misalignment. Yi praises Anthropic, one of the few labs still publishing actively, for constitutional classifiers. These look for topics or activations that light up during bad behavior instead of patching individual outputs. The hard part, both agree, is whose constitution it is. The host offers Grok's prompt, which reportedly tells it to align with Elon Musk's views. He then relays Anthropic's agentic misalignment study, in which leading models blackmailed people 80–96% of the time when facing replacement. Yi answers that few organizations run agents effectively in production. She argues that multi-agent hierarchies, where one agent directs others, need specialist design.
World models and multimodal risk. Yi points to Fei-Fei Li's World Labs, physics-informed video models such as Google's Veo 3, and Yann LeCun's JEPA. JEPA, she says, matched a crude bird drawing to a real photo without much context. Such models could reject a plan like vacuuming up spilled pasta by simulating it. But she warns that combining modalities also opens attacks across them, with text used to target image or video generation. The host imagines poisoning web text so a video model renders Xi Jinping as Winnie the Pooh. Yi says she has made an image of Biden be predicted as Trump, which is harder against frontier models' safety layers "but also yet not that hard."
How the attacks work. Yi says an attack runs in Python in a free Colab notebook against white-box or black-box models. To make a model see the host as Joe Biden, the attacker adds invisible pixel perturbations drawn from features the two share in vector space. She cites a Netflix paper questioning cosine similarity: embedding placement depends on arbitrary training factors, so Trump, Xi and Biden likely sit close together. Defense means gold-standard benchmarks tracked over time, which is harder for tasks such as translation where many outputs are correct and precision and recall don't suffice. Emerging attack vectors. The host describes prompt stealing as competitors extracting prompts worth "millions of dollars," and Yi notes jailbreaking can leak them. She describes slopsquatting, malware published under package names LLMs hallucinate, and Katherine Lee's DeepMind result that repeating one word made models emit PII. Cheap inference makes these "pretty trivial." She recommends SORRY-Bench for testing a model's susceptibility to coercion.
Funding women founders. Yi co-runs Generationship, which invests in early-stage female founders. She says about 2% of US venture capital goes to female founders, consistently over the years. The host says roughly 95% of early-stage Bay Area money goes to all-male teams.
Musa, a senior mechanical engineering student at Purdue, ends by describing his goal as building robots every day, ideally through full-time content creation and a startup of his own. His argument throughout is that hobbyists at home can build machines close to what Boston Dynamics makes, and his next step is a cheap, open-source mini version of his quadruped.
The rope-driven robot. His quadruped, CARA, uses conventional brushless motors and ODrive field-oriented-control controllers for compliant joints, but no gears. Grooved drums are coupled by tensioned rope anchored on both sides. Asked about slip and stretch, he cites the capstan equation: friction grows exponentially with each wrap. He says this was a hobby project done outside school, not coursework.
Origins and discipline. He took things apart as a child, tried to build a laser from a laptop at about 10, and found STEM through 3D printing in middle school. He chose mechanical engineering as the most versatile discipline and calls his software skills rudimentary.
Method. He iterates on the most unusual component first: about 20 actuator iterations, three leg prototypes, one robot. Printing versus buying depends on the project's intent and the time involved. He sees his work as showing hobbyists what is possible rather than pushing the state of the art.
What's next. CARA is blind apart from an IMU and is teleoperated. CARA Mini will be his senior design project, with software-literate teammates for vision. He is finishing a semester early. His tools are Bambu Lab printers, which he notes have sponsored his videos, and Autodesk Inventor, which he learned with help from a FIRST Robotics mentor.
*513 (* sends straight to voicemail in FreePBX, 513 being the station's extension); he used 10 seconds of ringing for the demo but suggests 20–30 in production.* triggers the relay. RFID unlocking worked with a key fob he had lying around; the bundled smartphone-app licences he did not test.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.
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.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.
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.
wget the Linux ARM64 installer, chmod +x it, run it with install under sudo, and allow about two minutes before browsing to port 11443.uos-server group to run its commands; a factory-reset U7 Pro XG then appeared ready to adopt.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.
a = 1, b = a + 1 and c = a + b, changing a to 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."set-inform manually, or paperclip-reset the device and re-adopt it.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."
The host stands up an early-access UniFi OS Server on a cloud-hosted Ubuntu droplet, secures it with a Let's Encrypt certificate and UFW firewall rules, and ends by adopting a remote Gateway Lite into it: a self-hosted, multi-site controller that, he argues, finally updates itself.
What it is. Ubiquiti aims the product at MSPs running gateways without onboard UniFi OS. He presents it as Ubiquiti's paid hosting released for self-hosting, whose chief gain over the legacy controller is UI-managed and automatic updates, plus Site Magic, InnerSpace and Identity. He steers anyone serious toward Linux over the Windows and Mac builds.
The server and install. On DigitalOcean he takes a $12-a-month droplet with 2 GB RAM over the $6 one that barely clears the 20 GB disk minimum, and points a Cloudflare A record at it. Podman and the x64 installer put version 4.2.23 in place in about 30 seconds, served on port 11443. The wizard's licence bars reselling hosting on the free software; he then enables multi-site management for per-customer sites.
Securing it. A certbot Cloudflare DNS plugin, a zone-scoped API token and a community-written import script load the certificate, with cron handling renewal; because the challenge is DNS-01, port 80 stays closed. Chrome kept flagging the site as not secure despite a valid certificate, which Firefox accepted. UFW then denies all incoming traffic except UniFi's listed ports and SSH, which in production he restricts to his WAN IP.
Adopting a site. Before the gateway connects he changes the default subnet, since sites joined by Site Magic cannot share one. The Gateway Lite, on a separate T-Mobile Home Internet line to mimic a client site, adopts into the hosted server and updates.
Benny Vasquez, chair of the AlmaLinux OS Foundation, ends up arguing that AlmaLinux's value lies in neutral, community-owned governance more than in being a free Red Hat copy. Since 2023 it has been a Red Hat–compatible sibling that can go beyond Red Hat, and she defines success as a project that outlasts everyone running it now.
From CentOS to Alma. Vasquez, a former CentOS user, recounts that CentOS ran almost 20 years as a free, community-supported rebuild, with about 40 million known active installs, but had no central governance. Red Hat hired its engineers, created CentOS Stream upstream of Red Hat, then dropped the downstream rebuild. She reads that as business logic: Stream gets earlier feedback, and the rebuild competed with the paid product.
No longer one-to-one. In 2023 Red Hat put its sources behind a login whose terms forbid building an operating system with them. AlmaLinux now builds from CentOS Stream while matching Red Hat's releases, so she no longer calls it one-to-one. The upside she claims is freedom: restoring RAID controller support dropped in 9 and pulling in upstream fixes sooner.
Governance and money. The foundation has a $250,000 annual budget and pays no one. Sponsors donate credits, hardware and engineers' time, and she says the project, not the sponsor, directs those engineers. It is a nonprofit so it cannot be sold. A member-elected board is balanced by a self-appointed engineering steering committee. She says some vendors refuse to support Alma even when large customers ask.
Hardware and new users. Red Hat 10 dropped x86-64-v2 chips. AlmaLinux committed to v2 for AlmaLinux 10's ten-year life and rebuilt EPEL for it, and she says adoption beat her expectations. She reports VFX studios running render farms and desktops on Alma, and a Linux desktop share of 5%.
Robert Osazuwa Ness, a researcher at Microsoft Research and author of the Manning book Causal AI, argues that causal AI is less a new algorithm than a discipline. The analyst writes down honest assumptions about how the data were generated, and existing probabilistic and deep-learning libraries handle the inference. He ends with a concrete workflow (question, DAG, identification, estimation, refutation). He also argues that LLMs are useful assistants at the formalizing steps but dangerous at estimation, and that the more interesting research direction is using causality to build better generative models.
Why AI ended up correlational. Ness says the book was prompted by a short passage in Judea Pearl's The Book of Why, in which a robot must work out why its owner is upset at being woken by vacuuming. He traces correlation-based AI to causal inference's roots in classical statistics, as in econometrics or epidemiology: separate the association due to a causal link from that due to a confounder. When deep learning arrived, it scaled that framework to more data and nonlinear, high-dimensional relationships without changing it. Cognitive science asks a different question: not what is true, but how to write algorithms that do what people's and animals' minds do. Humans, he says, reason well about intuitive physics and folk psychology.
The bridge to probabilistic programming. Ness wrote the book partly because practitioners were not linking DAG-based causal inference with probabilistic graphical models. Pearl contributed heavily to Bayesian networks, and reading a Bayesian network's edges as causal yields a causal model. The tooling ran from BUGS and JAGS to Stan, WebPPL, Gen and Pyro/NumPyro, and much of it was developed by cognitive scientists. Today's deep latent-variable models in Pyro can represent any 1990s Bayesian network. Ness argues causal inference had drifted from generative models because they handled latent variables poorly, and since a confounder is a latent variable, that obstacle is now gone. He notes that PyMC has added a do operator.
No free lunch. A causal model gains the ability to simulate an intervention, as a randomized trial would, only in exchange for stated assumptions about causal structure. Ness separates Bayesian priors, which encode the modeller's certainty about model elements, from causal assumptions about the process itself, though both inject assumptions to improve inference. Using machine learning's notion of inductive bias, he says causal theory tells you which questions a given set of data and assumptions can answer. Some cannot be answered with more data, only with more assumptions or more variables. A professor's line stuck with him: no statistical method removes bias from data.
A worked example and the tools. The book's running example is a gaming company asking whether side-quest engagement causes in-game purchases. Guild membership drives both, so the naive correlation misleads. A randomized test would force players into side quests, which is infeasible, and Ness adds ethics and cost as other reasons experiments fail. For tabular data he recommends DoWhy from the PyWhy suite. PyTorch and Pyro are for high-dimensional variables such as media, where a differentiable loss lets the library handle inference.
LLMs and causality. His paper on LLM causal reasoning, mostly testing GPT-4o, found strong performance at deciding whether A causes B without data, with the model acting as a causal knowledge base. It can propose DAGs, write DoWhy code and formalize questions such as the probability of necessity for a Netflix promotion. But its estimation advice was subtly wrong: correct only under strong assumptions it never stated. Ness finds causal generative models more interesting, such as one generative model per DAG node, or video-game generators that learn Minecraft's mechanics well enough to compose two games into a new one. He calls this work fresh.
The ladder and the workflow. Answering a listener, Ness explains Pearl's three rungs: association, intervention, and counterfactuals, which condition on what actually happened ("I didn't get vaccinated and got sick; would I have, had I been?"). Higher rungs need more assumptions, often mechanistic ones such as monotonicity, but never knowledge of the answer. His workflow: pose the question, encode the DAG, test identification, and if that fails, seek an instrument or a mediator for front-door analysis. Then choose an estimator on ordinary statistical grounds and finish with sensitivity analysis.
Brenda Darden Wilkerson, president and CEO of AnitaB.org, ends up holding that who succeeds in tech follows from access, exposure and the images people are shown, not from innate ability. She argues that AI and education should build on a person's own background and community rather than replace them.
Exposure. Wilkerson first met a computer in college. Computer science was not offered at her high school in an all-Black neighbourhood, which she attributes to policy, and two required programming classes in biomedical engineering were her way in. Hanselman calls it luck as preparation plus opportunity, and Wilkerson stresses the ripple effect on family and community.
Not bias, the crowd. Her nephew got the family Apple IIe in his room. Hanselman called this unconscious bias, and she pushed back: parents follow the images of who can, like the all-male Silicon Valley, or the ENIAC women once taken for models. They agree on "inertia."
AI and knowledge. She worries AI is being used to replace human creativity. She wants artists and medical people building tools for their own fields, and argues that because AI code is imperfect, knowing what should happen still matters. Hanselman adds that people feel their worth cheapened when worth means output. She replies that curricula should treat upbringing and community as the value.
Catching kids early. CS for All in Chicago Public Schools aimed to reach children before they are told they are bad at math. She says girls lead in math early, and that 35, the age she left engineering, is still the average age women leave tech.
Grace Hopper. She describes the celebration, with up to 155 countries and around 30,000 attendees, as physically myth-busting for women and men alike. Male teachers from her district returned seeing lopsided classrooms. Events run 4–7 November in Chicago and 2–4 December in Bangalore.