prev. engineer at @stainlessapi @stripe @google - now dad, husband, meat proxy 💍 @swetavajjhala
New Jersey, USA
- Tweets 4,506
- Following 1,682
- Followers 488
- Likes 5,235
the times are so wild that i thought this was real for a second
I knew that the Anthropic S1 would have some goofy risk factors but this has to take the cake
Community note
This post is satire.
sec.gov/Archives/edgar…
37
pour one out for the on-call teams over the christmas/new year's holiday
🚨BREAKING: It's over.
Cloudflare CEO Matthew Prince told Reuters he knows of additional severe vulnerabilities in other pieces of critical software that are currently being patched, beyond the PHP flaws that disrupted government websites this week.
Prince predicts busy months ahead as AI surfaces long-hidden weaknesses in the 'fabric of the internet': 'every single week it's going to feel like the sky is falling.'
56
I like how you can buy a high-end robot vac/mop combo, but you still have to pre-sweep and sometimes pre-mop just so it can run properly
2
85
you can tell ant/openai use agents heavily to build their products because there is so much complexity when you think about it... tui, desktop, mobile clients * cloud, local, ssh environments, all kinds of settings and modes, built-in diffing, browser, file ui, etc
1
35
now i will say, even with agents both of their products are super buggy and have a lot of weird edge cases. but there's no way that much functionality would be attempted without agents
i think it goes to show that engineers add complexity until they can just barely handle it
2
21
and the smarter an engineer, the more entangled crap they live with because they have the ability
it seems like they never just keep complexity low and go faster/smoother/more solid
7
it's wild that you can fire off a single prompt to a frontier model now and it gives you a chrome extension that uses AI to filter your X feed according to your preferences.
1
37
David Ackerman retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
886
2,634
19,850
4,563,701
okay i haven't really messed with muse or grok bot but dots is actually cool so far.
i asked it how i could communicate with it programmatically and it suggested making an mcp server. it generated one, hosted it, installed it, and linked me to it. i just oauth'd and it worked
2
313
I don't know why the codex remote control ability is so bad. Claude is just way smoother in this regard
1
63
I'm at the point where I am going to start a Claude session on the same machine just to have it watch the codex logs and tell me what's going on
1
37
rewatched this video and it reminded me just how horribly overengineered a lot of systems were before AI
Even with the risk of "AI slop", it's still harder to get away with this level of BS now piped.video/watch?v=y8OnoxKo…
34
Should have used an unpatched artifactory for communication
Last 3 days I did my first agent swarm. Wow these things are amazing, but hard to work with! It lasted about 36 hours before it fell apart and I had to stop everything. At peak there were 31 agents collaborating in 31 worktrees via a folder serving as a message board.
1
3
237