bstn retweeted
Got sick of Deutsche Bahn delays, so I built DelayBahn.
It uses historical train delay data to answer the questions that matter:
- How late is this train usually?
- Is there a more reliable alternative?
- Is that 7-min transfer actually realistic?
delaybahn.com
Many people are super scared of security issues of AI-coded apps, as if existing software was any gold standard. The thing to be afraid of is carelessness of the people behind it.
I'm a software engineer with an MSc in Computer Science and 25+ years in the industry. Here's why Coinkite's Coldcard bug isn't an oopsy, it's disqualifying.
Coldcard's seed generation silently fell back to a non-cryptographic "random" number generator for 5 years. Not because crypto is hard, but because Coinkite violated the most basic rule of secure system design: fail closed, never open. When a security-critical path can't be verified, the system refuses to run and throws an error.
It does not quietly substitute something weaker and carry on. This code should have refused to produce a seed. Instead it produced a predictable one and continue silently. ☠️
A company that lets its most critical code path go unverified for half a decade cannot be trusted with your keys.
Throw away your Coldcards. Never buy one again.
bstn retweeted
2 watches and a sports card around his neck lmao
Just goes to show you money can’t buy taste
Kevin O'Leary casually wearing a $1 MILLION+ Jordan card 😮💨
(via @LoganPaul)
This is probably the most unexpected pivot. I hope more AI firms explore possible crosstalk between domains - the unlocks could be huge.
Feels like the AI war has officially begun. Not by doing harm to others with AI but by restricting other nation’s access to it.
(the masquerade as national security issue is obviously bs)
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
bstn retweeted
Notes on 100+ Recent Technical Interviews
I interview a ton of engineers. Recruiting is the single most important technical CEO activity. Here are a bunch of impressions
1. There is a severe ZIRP engineering overhang that is currently washing out. They're getting laid off, managed out, etc. after having been massively overhired around 2020-2022. This is worst for Tier-2 big tech (think PayPal, Bill, etc.) but also FAANGs. These are overwhelmingly bad engineers.
2. This flood of unqualified but good-on-paper candidates makes this the hardest SF hiring market I have ever seen, due to the amount of nominally strong-looking candidates that you need to grind through.
3. I am highly skeptical of "AI as a cause for engineering layoffs". I think this is a large-scale polite fiction -- the companies don't want to admit they overhired, the engineers don't want to admit they are bad at their jobs. Everyone's blaming AI when it's really just the market rectifying itself.
4. Many of these engineers appear never to have had a real engineering function at their corporations. They're sitting in meetings, "making decisions about technology" but are unable to write software. I leave many interviews baffled by what exactly they were doing for so many years, let alone what their manager was doing.
5. I have interviewed some engineers from FAANG companies so shockingly nontechnical that I am forced to conclude that there is either (1) a lot of resume fraud going on or (2) that there are kickback grifts within those organizations -- people hiring their cousins and splitting the pay, that kind of thing. I have no other explanation.
6. There's a fun side-effect where after interviewing 20+ people from certain small but public companies, I actually feel like I am gaining a short sellers' advantage: there are financial technology companies out there that, knowing what I now know, I would never deposit a single dollar into.
8. Based on this "exhaust" data, and extrapolating a little bit, maybe aggressively so: I think folks like @pmarca are basically right when they say that ~every tech company is overstaffed by a factor of 2-4x. Whatever the reason -- staffing ahead of need, monopolizing certain engineer types (Google-style), headcount-driven promotion incentives, the reality is that a lot of these companies are not being run for the shareholders. The aggregate SBC expense is insane, and I expect this is going to get rectified eventually.
I'm sure that AI will play a role in rectifying this -- but I fear that people are going to blame AI for taking people's jobs when the reality is that the jobs were already long-gone, possibly always useless, but the highly-paid butts-in-seats remained. People will be mad at AI for taking away their lucrative sinecures. Maybe that's the same effect from a public policy perspective, but it feels different morally.
bstn retweeted
Don’t outsource your thinking.
Letting LLMs do your thinking is the new TikTok (or slot machine) addiction; intermittent reward, attention fragmenting across parallel streams you’re not really tracking.
State management is one of the hardest problems in agent harnesses, and the moment you’re running more than three terminals at once, you’ve stopped managing agents and you’re just nudging them along while losing track of what they’re actually doing.
I’d even cut it to one session. Make yourself sit with the problem before you hand it off. The reading, thinking and planning muscles returning in the first week will cause some discomfort.
Slow down. Do the deep work.
I'm 22 years old and Claude Code is deteriorating my brain.
Every single day for the last 6 months I've had 6 to 8 Claude Code terminals open, waiting for a response just so I can hit 'enter' 75% of the time. And it's doing something to me.
In convos with a couple of friends, it's been a point that's been brought up pretty frequently.
None of us feel as sharp as we used to.
I don't know if it's just us, or others in their 20s are feeling the same thing, but it's something I've been thinking about a lot.
P.S. I know this is a problem with my reliability/usage of it, not Claude Code itself, but the effects are real nonetheless
Not even three weeks after Figma MCP promo…
Everyone’s mad at Anthropic for cutting OpenClaw off subscription plans. Let me offer a different take.
OpenClaw’s prompt caching had ~10% hit rates because the system prompt was a monolithic blob mixing static and dynamic content.
A single “what model are you?” burned 9,600+ tokens across 4-5 hidden background API calls.
One instance could consume $1,000-5,000/day in compute on a $200/month plan.
I spent days making my session context management compatible with prompt caching. The result: 80-90% savings from the second turn onward. It’s hard. It’s limiting. But that’s the engineering work.
The outrage isn’t about Anthropic being evil. It’s about a product that got glorified because nobody had to pay the bill personally. The moment the subsidy disappears, the architecture gets exposed.
If your tool isn’t sustainable without someone else eating your compute costs, it’s a technical debt problem, plain and simple.
The OpenClaw team clearly has the engineering talent to fix this. The caching infrastructure is there, they just need to make it actually work at scale. If they do, the product stands on its own merit. And that’s a much better foundation than a subsidized free ride ever was.
Ars contexta basically @arscontexta
Wow, this tweet went very viral!
I wanted share a possibly slightly improved version of the tweet in an "idea file". The idea of the idea file is that in this era of LLM agents, there is less of a point/need of sharing the specific code/app, you just share the idea, then the other person's agent customizes & builds it for your specific needs.
So here's the idea in a gist format: gist.github.com/karpathy/442…
You can give this to your agent and it can build you your own LLM wiki and guide you on how to use it etc. It's intentionally kept a little bit abstract/vague because there are so many directions to take this in. And ofc, people can adjust the idea or contribute their own in the Discussion which is cool.
