Gemini 4 Argon day. Woodward's demo trio (529/414/209): 'Gemini 4 Argon: At 96 millimeters wide, the hardest part of manufacturing this custom furniture leg was deciding how many stacked spheres to give it'; 'The latest Gemini model made this'; and the 24-word-prompt synthesizer that turns galloping horses, rippling water, and your own hands into music. Levie's Box eval (376): 'At Box, we've been testing Gemini 4 Argon on our Box AI Complex Work eval for the past week. It's a huge upgrade over Gemini 3.1 Pro, jumping by 14 percentage points across all our tests, and firmly at the frontier' — with per-industry wins: financial services +18 (funding-gap analysis built exactly as specified, every line broken out), healthcare +25 (the only model that went back and pulled the offending drugs from the ICU regimen so the plan matched its own diagnosis), technology +37 (pulled the right month's user/retention figures and correctly flagged North America, where 3.1 Pro reported retention above 100% and pointed at the wrong region), energy +16 (consistently caught the smart-meter log stopping in November), consumer products +33 (saw existing inventory could bridge an early capacity gap, making the smaller expansion viable). 'Gemini 4 is far more capable on math, logic, analytics, and deep knowledge of a variety of business domains. It also works far harder to complete tasks fully without stopping short.' Shipper's counterpoint (235): 'gemini 4 argon really struggles on our senior engineer benchmark task — comes in at 21% far below Sol 6 and even Muse.' And Rauchg's congrats (1841): 'Congrats Gemini team. Never bet against Google, the company that brought us Transformers and has overwhelming data, engineering talent, and compute. Everyone has, rightfully, pivoted back to coding. Coding agents are the killer app of AI. We're firmly at an elysium for developers and builders. Incidentally, getting Muse Wonder and Gemini Argon back-to-back is a reminder why you as an individual, and your company, should stay model agnostic and gateway-maxxx / fx-maxxx.'

Sottiaux's Day 9 (10414 likes, top of the day): 'We've rolled out optimizations to how we store state on disk for the chatgpt & codex desktop app that should reduce the amount stored on disk by 60%. Small ship but we've heard enough feedback about power users running out of disk space.' Swyx's Kill My SaaS results (86): 'A lot of people talk about the death of SaaS but mostly in theoretical terms. We took a real $43,000/year SaaS we use and launched a global hackathon to clone it in a weekend, with both AI and REAL HUMAN evals, using any models any harness. 72 submissions, a lot of manual evals and ~100B tokens later, here are the results… what's still hard about killing nontrivial, mid-tier SaaS, and the winners of the $10,000 Kill My SaaS contest with qualitative interviews!' Garry Tan's unlock (327): 'Giving your harnesses access to a real browser with your real credentials is a real unlock - highly recommend @AsideAI.' Amjad's triple: MXC.exe (191); 'brb, adding puppies to all my sites' (167); and the diffusion take (134): 'You can pause frontier research today, and it would take AI another 10 years to diffuse into the economy. We're working with solutions partners to accelerate that.'

Levie's cyber thesis (60): 'Cyber is going to be one of the biggest spaces for AI agents in the future. Obviously the risks of software exploits and the attack surface area increases 100 fold due to the ability for agents to work through any crack to find what they're looking for. But this goes the other way as well. The ability to have agent swarms review every line of code, test any number of permutations of exploits, and research any incident faster than humans as they happen, will be a huge boost… You can never hire enough people or work through the backlog of all known risks. Agents actually give you a chance finally to have far better coverage' — with plays across labs, data providers, network/firewall, identity, endpoints, SDLC, and data protection. Thariq's homecoming (858): 'excited to be back home.' Nikunj's Jake appreciation (50): 'All my priors get shattered when I chat with @JustJake every month… @Railway is an n of 1 company because Jake runs it HIS way.' Peter Yang's trio: the Bay Area homeowner special (101): 'Owning a house in the Bay Area is fun because you can just randomly spend five-figures on yard work'; 'Poteto cracks me up with her trolling' (54); and the not-so-hidden recorder (28): 'Is it just me or is an usb recording device plugged into a conference room tv "concealed" in toilet paper not a very good disguise at all.' Plus Nan Yu's 'Told y'all' (78), Garry Tan's SF politics post (153), Swyx's talent-density note (35), and Zara's video praise (33).

A brand-new Training Data episode (new to this digest) with David Silver and Lasse Espeholt of Ineffable Intelligence — the AlphaGo/AlphaZero legends' new venture building 'a superlearner for the age of experience.' Core threads: the superlearner defined — something that discovers completely autonomously, applicable to embodied domains, where LLMs' human-in-the-loop effort and human data may actually be 'weighing things down' ('thinking about these things in terms of language might not even be the right paradigm'); crucibles — simulated environments with enough emergent complexity (physics engines with millions of interacting objects, millions of interacting agents; 'Go is just a 19x19 grid — single-agent Go would be the most boring game imaginable; the moment you put a second agent in, you get an infinite onion of layers of complexity') for agents to rediscover everything from motor control to language to math to science; simple rewards instead of breadcrumb trails — 'we're not pulling the agent by its nose'; evolution optimized one fitness signal and gave rise to all of life's complexity, humans pursuing simple goals cascaded from coordination to communication to trading to arithmetic to tools to science; why AlphaZero took just hours to beat two years of human-in-the-loop AlphaGo work ('it has to be an active choice to let go of the human data'); human language is not the endpoint of reasoning ('maybe using parallelizable, rich, complex neural reasoning will be far more expressive'; using 1000-year-old language would bake in flat-world beliefs); weak environment-agnostic priors are fine, trillion-item datasets are not ('we're not religious about it — it's a pragmatic position'); timing — Silver the super-forecaster (called Computer Go's fall to the year, ten years out; called StarCraft and the IMO) says this is the right time, with 100,000x more compute available than OpenAI's hide-and-seek era; the cluster — first customer of the largest Rubin cluster on GCP by year-end, 'one rack is basically one machine' with the fast iteration RL needs, everything kept on GPUs (LLM companies now rediscovering the CPU-sandbox problems he solved years ago); alignment — language-based alignment is reaching its limits as agents 'learn to say one thing and act in a different way'; the most important thing Ineffable can do is build the science of superintelligence in simulation — what conditions make a society of superintelligences interact peacefully with lesser intelligences — and gate real-world action on that evidence; monitoring beyond language (shadow actions; superintelligent monitors regularizing the magnitude of other agents' actions); pace the real world (supportive of the Demis/Dario direction) but accelerate in simulation; the 20-person Avengers-grade team (four cofounders from AlphaStar; 'I've had to say no to a lot of people I've worked with — our bar is so super high') and its research-led culture; and the forecast: strong evidence within a year, true superintelligence in three ('my timelines have accelerated because of coding agents — buckle up').