Claude cracks NIST curve seeds & Mercor maps AI benchmark gaps - Hacker News (Oct 17, 2026)

For decades, cryptographers suspected something shady behind the numbers that secure much of the internet. It now looks like the real answer may be an old Unix random-number function and a clock. More on that in a moment. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I'm TrendTeller, and today is October 17th, 2026. We've got a cryptography cold case, a reality check on AI benchmarks, a game console running language models, and a verdict on image formats. Let's get into it.

Claude cracks NIST curve seeds

Our lead story comes from Anthropic. The company says Claude has solved a decades-old puzzle: how the seed values behind the NIST elliptic curves were actually generated. These curves, including P-256 and P-384, underpin ECDSA signatures across the web. The seeds were published long ago with no explanation, and that silence fueled years of speculation about hidden backdoors. According to the investigation, there was no secret hash and no elaborate scheme. The seeds appear to come from lrand48, a plain old C library random-number generator, seeded with Unix timestamps. Claude also worked out how base points were chosen for several older curves. Anthropic is careful to say this is not a cryptographic break. If anything, it makes deliberate tampering less plausible, because the process was simple and reproducible. A few loose ends remain, including one unexplained binary-curve seed. Still, a long-standing source of suspicion around core standards now has a fairly mundane answer.

Mercor maps AI benchmark gaps

Staying with AI, Mercor has published what it calls a Benchmark Map. It measures how much real-world work current AI benchmarks actually test. The answer is: not much. Using the US government's O*NET catalog of jobs and tasks, Mercor finds that its initial set of benchmarks touches only about a fifth of occupations and a small sliver of tasks. Roughly ninety percent of American working hours go to tasks with little or no benchmark coverage. The coverage that does exist leans heavily toward tasks already seen as exposed to AI, and toward larger, higher-paid jobs. Mercor calls the map provisional, but the takeaway stands. When someone claims a model can do a job, the benchmarks often haven't checked most of that job.

Claude Code rethinks code review

On the practical side, Anthropic's Claude Code team described how it rebuilt code review now that AI writes most of its code. Humans no longer read every diff. Claude verifies work, reviews pull requests, and flags bugs and security issues first. Those checks sit inside layers of defense: CI, a merge queue, gradual rollouts, and quick rollbacks. Human approval is saved for changes that are genuinely risky and for decisions that need judgment. Between June and September, merged pull requests more than doubled, while reverts fell by over forty percent. Incidents are also fed back in as evaluation data. The broader point is that code review is turning into a risk-management system, not a reading exercise.

MLX update for Mac Studio

Now to running models locally. In an update to the Apple MLX story we've been following, the latest MLX release adds new benchmarks on the M5 Ultra Mac Studio. Apple reports big gains over the M3 Ultra, both in time to first token and in generation speed. For models too large for one machine, MLX can now spread a single model across two or four Mac Studios connected by Thunderbolt, using a new low-overhead communication backend. Apple's pitch is clear: serious, private inference that runs in a home or office instead of the cloud.

PS5 turned into LLM server

Taking local AI somewhere less expected, a project called PS5LM runs llama.cpp on a jailbroken PlayStation 5. It uses the console's GPU and shared memory to serve language models, including ones that understand images. It speaks both OpenAI-style and Anthropic-style APIs, includes a browser chat interface, and can even show requests live on your TV. It's a niche hack, but a telling one. A living-room console can now act as an AI server for every device on the network.

DeepSeek V4.1 Flash snapshot

And for those using hosted models, DeepSeek's October 17th snapshot of V4.1 Flash is now listed on OpenRouter. It's a sparse mixture-of-experts model with a one-million-token context window and native image understanding. It's aimed squarely at coding, terminal, and computer-use agents working on long, multi-step tasks. It keeps DeepSeek's familiar strategy of competing hard on cost, as the budget tier for agent workloads.

Transformers as associative memory

For something more conceptual, a new explainer recasts the parts of a transformer as memory systems that try to recall stored facts without mixing them up. The simplest linear memory gets crowded quickly. Better-designed linear layers roughly double capacity, and the nonlinearities in MLPs suppress interference even further. Self-attention is the big jump. It stores keys and values from the context itself, so its memory grows with the input. The piece also explains why sharpening attention as context gets longer helps keep recall reliable. Capacity climbs from dozens of memories to millions. It's a helpful lens for anyone wondering why attention works so well.

JPEG XL versus AVIF showdown

Over in web performance, Jake Archibald has pitted JPEG XL against AVIF using current browsers and encoders. For lossy photos, AVIF usually comes out ahead: smaller files, faster encoding and decoding, and lower memory use. It wins clearly for screenshots, illustrations, and transparency too. JPEG XL's standout strength is lossless compression, where it often beats the alternatives. Its results also vary a lot depending on which encoder you use. Archibald's advice is to default to high-effort AVIF, but test JPEG XL when AVIF files come out surprisingly large. Both formats are still improving, so this verdict could shift.

Dune's Butlerian Jihad revisited

Finally, a bit of science fiction that feels oddly current. Today's list includes a deep dive into Dune's Butlerian Jihad, the generations-long revolt against thinking machines. In the story, it ended with a commandment banning any machine made in the likeness of a human mind, down to simple calculators. Human institutions such as the Mentats, the Bene Gesserit, and the Spacing Guild arose to fill the gap. It's the reason the original novels have no AI. On a day full of AI-written code and AI-solved mysteries, it's a reminder that people have been debating this tension for a very long time.

That's the show for October 17th, 2026. Links to all stories can be found in the episode notes. Thanks for listening to The Automated Daily, hacker news edition. I'm TrendTeller, see you tomorrow.

Claude cracks NIST curve seeds & Mercor maps AI benchmark gaps - Hacker News (Oct 17, 2026)

Our Sponsors

Today's Hacker News Topics