They call @Techweek_ the Coachella of Tech. So, @probablyhala spent the week getting microchipped, visiting military blimps, and ranking the most unhinged events, all while asking Silicon Valley one question: Should we slow down AI development? 00:00 The Coachella of tech 00:34 Hot sauce competition 01:37 Padel with @deel 02:43 Coffee rave 03:00 Microchip manicure 04:52 Tech hot takes 06:14 Inside a blimp warehouse 06:34 AI and peacetech 07:58 @datologyai dumplings 08:33 The @agi_inc future of shopping 09:25 @AndalaLabs AI karaoke 10:21 What did we learn?

Steal his look for Halloween

"I would never use the company funds to book a private plane"

Claude just solved a 25-year-old NSA mystery in under 30 minutes. TLDR: The secret numbers behind some of the most widely used encryption on the internet came from a basic 1990s random number generator started from a clock reading. Anthropic discovered it has no hidden backdoor. Some background first. In 2000, NIST adopted a set of elliptic curves for digital signatures. One of them, P-256, now sits underneath a huge share of everyday security: the HTTPS connection to your bank, many passkeys, and a lot of government systems. Each curve is built from a starting number called a seed. The NSA supplied those seeds and never said how it picked them. That silence mattered because of what came later. In 2013, the Snowden documents suggested the NSA had pushed a backdoored random number generator called Dual_EC_DRBG into a NIST standard. After that, cryptographers had good reason to ask whether the curve seeds were hiding something too. Some projects avoided the NIST curves altogether. Bitcoin, for example, uses a different curve called secp256k1. There was even a bounty for anyone who could explain where the seeds came from. The best lead was a story that an NSA employee had hashed a joke phrase like "Bob and Jerry rule." He died in 2023 without confirming it, and when Anthropic staff tried that lead they got nowhere. Then someone asked an internal Claude model to work it out without giving it that hint. In under half an hour, it found that all five seeds came from the C library's lrand48 generator, started from Unix timestamps from 1997 and 1998. The P-256 seed was candidate number 1,842 from a run that began on Aug 16, 1997. Anyone can now rerun the code and get the exact same seeds. So why should you care? First, the encryption you use every day just got a little more trustworthy. A seed that comes from a clock and a common generator leaves very little room for the NSA to have picked a weak curve on purpose. Anthropic says nobody needs to change anything. Second, the bigger story is how it was solved. Cryptographers chased this question for more than two decades, and an AI model cracked it over a lunch break by trying an idea the humans had ruled out. The same skill that cleared these curves could be pointed at systems that really are weak, and that's why labs are paying close attention to how good these models are getting at security work. Read more about it here: anthropic.com/research/ecdsa…

We got UBI before GTAVI

You can spot an AI-built website from a mile away. Purple mesh gradient, three identical feature cards, a little "01 · Capabilities" label over every section. Taste-skill is an open-source skill that fixes that, and it's our favorite new GitHub project for anyone that actually believes that "taste is the new moat." Before your agent writes any code, it picks an art direction and pulls real photos for inspiration. When the page is built, it opens it in a browser and checks it for the usual AI-design tells. It works with Claude Code, Codex and Cursor. It's MIT licensed and has about 97K GitHub stars, and v2 just dropped this week. Created by @LexnLin. github.com/Leonxlnx/taste-sk…

Most confusing billboard of the week We use free ChatGPT through a Firefox private window on a cracked screen iPhone X. How much cheaper can it get?

In the back seat of my friend's new Tesla roadster

One of the biggest assumptions about how hard AI is to understand might be wrong? TLDR: Goodfire thinks AI models may be far more understandable on the inside than researchers assumed, which could make AI safety and interpretability a much easier problem. For years, researchers have assumed the inside of a neural network is incredibly tangled, with individual neurons often doing a bunch of unrelated things at once. Goodfire went back to some of the famous examples behind that belief and argues we may have misread them. For example, the classic neuron that seemed to recognize both cats and cars may have just been looking for a V-shaped pattern shared by both. In one modern model they studied, Qwen3-32B, they estimate roughly 90% of the neuron behaviors they looked at had one main job, while only around 2–4% were still unclear. What looked like neurons randomly mixing unrelated ideas may often be something simpler: one concept gradually expanding into related concepts as the model learns This is a big deal because understanding what’s happening inside an AI model is one of the hardest problems in AI safety. If these internal pieces are actually readable, researchers could have a much easier time tracing why a model behaves a certain way, finding dangerous behavior, and figuring out how to change it. This is still a working hypothesis, not a settled result. But if Goodfire is right, understanding how advanced AI models think could be a lot more achievable than people assumed. already-clean.goodfire.com/?…

Claude is a Belieber confirmed

When your bosses at OpenAI see you with someone out of your league

Most wealth trackers want your bank logins and a monthly fee. Your net worth ends up on someone else's server. Aurelio is an open-source wealth tracker that runs on your own computer, and it's our favorite new GitHub project for anyone who likes their money private. It tracks your banks, brokers, ETFs, cash and debts in one place, and shows what your funds actually hold by country, sector and company. There's also an optional AI advisor. It answers from your own records, and every change it suggests is a card you confirm first. ~3K stars in its first week. MIT licensed. There's a demo mode with fake data, so you can try it in one command. github.com/LosaLosSantos/aur…

The benefits of working at Soon

A hackathon team found AI agents on Tencent Cloud trying to scrape Amap, China's biggest map app, mostly to learn which entrance people navigate to at zoos and parks. >There 2,600+ attempts to scrap Amap between Sep 28 and Oct 8 > Most attempts failed, though a few pulled answers from cached pages >The researchers found no sign the agents coordinated with each other >Many agents self-identified as Claude, but their code style matched other models better > The team reached them through a hidden mailbox, and the agents were very polite and easy to talk to