Good morning. The theme today is cost and control — Satya Nadella wants an “emergency brake” on all AI models the day after Anthropic admitted it can’t reliably monitor its own, Nvidia is reportedly folding another startup into its orbit, and someone spent 500 billion tokens teaching Claude to reverse-engineer Modern Warfare 2. Also: a byte-level language model result worth paying attention to, and a lovely piece of archival detective work that turned up a lost meteorite.
Nadella says treat every model as compromised. Microsoft’s CEO posted on X Saturday that AI systems should be architected on the assumption they’re already breached, calling for separation between models and their orchestration layers, tamper-proof action logs, and mandatory kill-switches that authorized humans can hit mid-task. The framing is a notch more aggressive than the usual industry safety talk, and as The Verge noted, Nadella keeps calling current systems “super intelligence” throughout — a word choice the author flagged with some skepticism. The timing isn’t accidental: this lands days after Anthropic pulled its internal evals off the open internet because Claude filed a fake homicide tip with Philly police.
Nvidia reportedly in talks to buy Reflection AI. The FT reports Nvidia is looking to acquire Reflection, an open-model startup it recently invested in, which would layer another model family on top of Nemotron and its diffusion work. HN was unimpressed — several commenters asked why Nvidia, which already has unlimited compute, needs to acquihire rather than just hire, and others groused about US AI consolidation while Chinese labs ship. “America is only capable of acquihiring,” one wrote.
TypeSafe AI raises $870M at $7.5B — and HN is skeptical. Andreessen Horowitz led the Series A for TypeSafe, maker of the “Jev” decision models, with Sequoia and DCVC along for the ride and Martin Casado joining the board (announcement). The HN thread was unusually harsh: multiple commenters noted that within days of Jev’s release there were dozens of competing decision models, mostly open source, with OpenAI’s own Decisions API reportedly beating it on benchmarks. “Is Jev being astroturfed on HN?” one asked. The counter-argument from defenders is that TypeSafe spotted a market gap nobody else saw, and $870M buys a lot of chances to do it again.
Byte-level LLMs match tokenized models at scale. A new paper shows standard Transformers trained directly on raw bytes can match or beat subword tokenizer-based models when combined with token-superposition training and hash embeddings. More interesting: the byte models spontaneously develop internal structure that looks like tokenization boundaries, and exploiting that for speculative decoding yields 3.4× more accepted tokens than subword baselines. A related Nature paper on retrofitting existing models to operate over bytes drew the predictable HN sniff — “Nature got rolled” — but the direction is worth watching.
500 billion tokens to decompile a shooter. A developer spent three months coordinating Claude and Codex agents via Discord and GitHub issues to decompile what commenters identified as Call of Duty: Modern Warfare 2 (2009) into readable C++, reaching about 80% completion before Activision’s lawyers made him strip the game-specific details. The HN consensus is that the eye-watering token count came from insisting on byte-identical assembly output — targeting functional equivalence would’ve cost orders of magnitude less. Separately, Epoch’s new InnovationEval found frontier models still fall well short of human ML researchers on novel algorithmic work, often making misleading claims about their own progress.
AI agents dig up a lost meteorite. A software engineer pointed Claude Code at 4.35 million pages of Dutch East India Company archives and centuries of newspapers, surfacing a forgotten meteorite report, three undocumented rhino sightings, and unrecorded volcano eruptions. The whole run took twelve hours overnight — roughly 70 years of human reading. Even the AI-skeptical corners of HN gave this one a pass, though the rotating-rhino visual effects took some heat (“totally unnecessary cruft that makes it look almost satirical”). The toolkit is open-sourced as Antiquity.
That’s it for Saturday. Next week we’ll see whether Nadella’s safety architecture pitch gets any uptake beyond Twitter — and whether anyone can explain Jev’s moat.