Discovery Loop — 888 points
Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals — four of the most cited researchers in AI and distributed systems — have launched Discovery Loop, an independent public benefit corporation focused on automating the entire experimental loop in science and engineering. The pitch: propose, run, and learn from experiments in parallel at massive scale, compressing iteration cycles that currently bottleneck scientific progress.
They’re starting with ML research as their own first customer, then expanding outward. The grand ambition is tackling NAE Grand Challenges — better medicines, clean water, solar energy. The team pedigree is undeniable, but “automate all of science” is the kind of mission statement that sounds incredible in a pitch deck and takes 20 years to materialize in practice. The real test will be whether their automated ML loops produce genuinely novel insights or just hyperparameter tuning at scale.
This is the same Jeff Dean departing Google (see next story), so there’s a neat narrative thread here. Google is reportedly a founding investor and Cloud partner, which means Discovery Loop gets infrastructure without the corporate overhead. Whether a “public benefit corporation” structure actually keeps them honest long-term remains to be seen.
Changes at Google DeepMind: Demis Hassabis moves to Chair, Jeff Dean departs — 807 points
Sundar Pichai announced a major leadership restructuring at Google DeepMind. Demis Hassabis is moving from CEO of GDM to Chair of GDM and Chief Scientist of Alphabet — a role that lets him focus on “shaping the future of AGI” and his work at Isomorphic Labs (drug discovery). Koray Kavukcuoglu, a 13-year DeepMind veteran and current CTO, steps up as SVP of Google DeepMind.
The bigger news buried in the announcement: Jeff Dean is leaving Google after 27 years. He and Sanjay Ghemawat — the pair behind Mapreduce, the Google File System, and arguably the neural network revolution at Google — are launching an independent venture. That venture turns out to be Discovery Loop (above).
The timing is telling. Google is simultaneously claiming AI momentum (Gemini at 950M+ monthly users, Gemma models at 900M+ downloads) while its most legendary engineers walk out the door to build something new. Hassabis moving to a strategic “Chief Scientist” role reads as either a visionary repositioning or a polite step back from operational control. The “foothills of the singularity” language is pure Hassabis — grandiose but sincere. Whether Koray can maintain GDM’s research edge while also shipping competitive products is the open question.
Mario Meets Pareto — 737 points
Antoine Mayerowitz built an interactive tool that applies Pareto optimization to Mario Kart 8 builds. The core insight: with tens of options for drivers, bodies, tires, and gliders — each with distinct stats for speed, acceleration, handling, weight, offroad, and mini turbo — the combinatorial space explodes. Most choices are dominated (strictly worse on all dimensions), and Pareto’s century-old concept of efficiency cleanly eliminates them.
The interactive visualization lets you toggle between stats and watch the Pareto front update in real time. It’s a genuinely effective teaching tool for multi-objective optimization — you immediately see why Koopa is always a bad choice and which builds occupy the efficient frontier. The final decision still depends on your play style, but Pareto at least guarantees you’re not leaving performance on the table.
This is the kind of “serious math applied to silly domain” post that HN loves, and for good reason. The underlying principle — that most options in any decision space are dominated and can be eliminated without loss — applies to everything from hiring to cloud architecture. The Mario Kart framing just makes it visceral.
Crime Pays but Botany Doesn’t — 589 points
The reading list from Crime Pays but Botany Doesn’t — a self-taught botanist’s guide to learning plant taxonomy and identification from scratch. The author’s voice is profane, passionate, and refreshingly anti-academic: “Don’t be intimidated. With the internet, you have 24 hour access to the library. Use it.”
The core advice starts with understanding Latin terminology and why Linnaeus’s system persists despite its historical baggage. Common names don’t scale — “cedar” refers to eight unrelated plants, while Cedrus is precise. The author takes a swipe at “performative leftist nitwits” who attack taxonomy itself, arguing you can acknowledge Linnaeus’s flawed ideology without discarding the functional system he created.
This resonated on HN because it’s fundamentally about self-education in a field gatekept by academia. The botany angle is niche, but the pattern — domain with intimidating jargon, passionate outsider providing an on-ramp, tension between institutional knowledge and autodidact energy — maps cleanly onto programming, ML, and dozens of other fields.
Beating GPT-5.6 Sol on retrieval with 100x cheaper open models — 414 points
Neon and Castform claim a 4B parameter open-source model, post-trained with Castform, matched GPT-5.6 Sol’s retrieval accuracy while costing 100x less. The approach: instead of one-shot embedding search (the 2022 paradigm), they use agentic retrieval — the model plans and executes multiple search iterations in a loop, refining queries against Neon’s Postgres-based search extensions.
The cost argument is compelling on its surface: if a 4B model can match frontier retrieval with multi-hop search, the economics of RAG pipelines shift dramatically. But “matched on retrieval” is doing heavy lifting here — retrieval accuracy on a benchmark isn’t the same as end-to-end task performance. The real question is whether this approach degrades gracefully on messy, real-world queries where the search space isn’t clean.
The broader trend is clear: the moat around frontier models for specific tasks is eroding fast. Post-training on domain-specific data with smart tool use is proving more effective than raw model scale for retrieval-heavy workloads. If this holds, the winners shift from “who has the biggest model” to “who has the best training data and tooling integration.”
How to Make a Nintendo 64 Game in 2026 — 406 points
Dominic Szablewski documented the making of Xibalba 64, a Wolfenstein 3D-style FPS released as a physical cartridge for the N64 via Modretro’s M64 clone. This is reportedly only the second physical N64 game release since the console’s commercial life ended.
The technical journey started with Impact, a JavaScript game engine Szablewski built in 2010, which he rewrote in C “for no reason” two years ago. That C port (high_impact) has modular platform and rendering backends, making it viable for N64 development. The N64’s Reality Coprocessor — with its fixed-function RDP and programmable RSP — required building on top of libdragon (the open-source alternative to Nintendo’s proprietary libultra) to avoid copyright issues.
The post is a love letter to constrained hardware programming. The N64’s quirks — big-endian MIPS, the RSP’s vector processing, the RDP’s idiosyncratic instruction set — demand a level of low-level understanding that modern game development has largely abandoned. The fact that someone is shipping physical N64 cartridges in 2026 is both absurd and admirable.
The Title Cards in Blade Runner Are Amazing — 402 points
Michael Lopp (Rands in Repose) wrote an appreciation of Blade Runner’s typography, triggered by rewatching the film on a flight while researching typefaces for his terminal setup. The punchline: the entire title sequence uses a single typeface — Goudy Oldstyle — deployed with variations in size, weight, and case to create distinct emotional registers.
The post meanders through fixed-width coding typefaces (SF Mono, JetBrains Mono, etc.) before arriving at Blade Runner, which is either a structural flaw or exactly how an attention-span-challenged tech executive’s brain works. The core observation is solid: great typography is invisible until you look for it, and Goudy Oldstyle’s warmth in the title cards creates a noir intimacy that a more “futuristic” typeface would have fumbled.
This is pure vibes content — no technical depth, no actionable insight — but it landed on HN because typography appreciation is one of those topics where everyone has opinions and nobody writes about it well. Lopp does.
Born Against: Why Hobby Programming Communities Resist LLMs — 388 points
Fogus explores why niche hobby programming communities — chess engine developers, OSDev, demoscene, code golfers, language designers — are increasingly hostile toward LLM usage. The thesis: in these communities, the process of mastering a difficult field is the product. A working program is a nice-to-have. Using an LLM to generate the output misses the point entirely.
Respect in these circles is earned through years of sharing elegant code, demonstrating deep domain knowledge, and displaying genuine curiosity. The community doesn’t care if your code works — they care that you know why and how it works. LLM practitioners who burst in with generated solutions are perceived as the Kool-Aid Man crashing a meditation retreat.
The nuance Fogus adds: even when early LLM engagement was earnest, the well was poisoned by practitioners lacking deep understanding and a vitriolic subset viewing LLMs as cheating. His framing of LLMs as “force multiplier, not surrogate” is the most balanced take — in expert hands, they’re a lever; in novice hands, they rob you of the craft. The implicit counterpoint: most people using LLMs for code aren’t in these communities, and the gatekeeping that preserves craft also preserves insularity.
Humans Missed 1 in 3 Threats Approving AI Agent Commands — 207 points
Alex Wauters analyzed 40,000 game runs of his browser game where players act as human-in-the-loop reviewers for an AI coding agent’s commands. The results: average players missed 1 in 3 threats (66.3% accuracy), 32.9% of sessions ended with negative scores, and the single most-missed command was npm run analyze — approved 64.7% of the time despite containing an exfiltration payload.
The npm run blind spot is the real finding. The command itself looks innocent, but it executes whatever script is defined in package.json — which could be injected malicious code. Players missed npm-based attacks 52.5% of the time versus 28.4% for other exfiltration methods. Hiding payloads behind familiar script names roughly doubles success rates, even when the payload is visible in the agent’s history log.
The data also confirms permission fatigue: miss rates improve briefly (warming up) then climb toward the end of sessions. The over-blocking problem is equally damning — benign commands like npm config set registry were blocked 59% of the time, creating noise that erodes vigilance. The takeaway is bleak: command-level approval as a security model is fundamentally broken because users lack context about what has changed in the codebase. Sandbox everything, separate your credentials, and stop pretending human review at the command level is a meaningful safeguard.
The Throughline
Today’s front page tells a story about the AI industry’s talent redistribution. Jeff Dean and Sanjay Ghemawat leaving Google to build Discovery Loop, Demis Hassabis stepping back from operational DeepMind leadership to focus on AGI strategy — these aren’t small moves. The people who built the infrastructure of modern AI are betting that the next chapter happens outside big tech’s operational machinery.
Meanwhile, the tools are getting cheaper (a 4B model matching GPT-5.6 Sol on retrieval), the security model is failing (humans can’t reliably review AI agent commands), and the communities that value craft over output are pushing back against the LLM wave. The tension between “AI as force multiplier” and “AI as replacement for understanding” is the defining debate of the moment — whether you’re a hobbyist chess engine developer, a DevOps engineer approving agent commands, or a scientist whose experimental loops are about to be automated.