People need to chill. This is likely just because of embeddings (vector) search, not an evil/genius idea.

I hope not all GCCs go this way, but I fear that a lot of them will 😔

Apple showed how llms can be made really useful in human chat apps, even if it is in only small delightful ways. Meta completely dropped the ball here, treating WhatsApp only as a distribution channel to compete with other model labs. They are not thinking of the users at all.

Good take on Anthropic’s automatic fallback from Fable 5 to Opus 4.8 and more importantly silently degrading its performance.

There's some wild disconnect that people believe productivity is spending 6-figures on inference in a month. You know what happens if that remains true? We hire more software engineers and spend less on inference. "Loops" - or any other equiv - only work if compute is cheap.

A friend’s wife need A negative blood in Malad. Any leads? Please spread the word.

So good. “Get feedback early” feels obvious, but it’s a totally different thing to understand it viscerally. I thought we do this as a habit at udaan (not just engg, even biz folks). But as I read this, I felt that there is so much more we could do.

An internal Slack post I shared with the Amp team a few weeks ago: Building software is learning. registerspill.thorstenball.c…

using AI for coding is a deeply technical engineering craft most people don't approach it as so, and don't get the results we associate with high craft but the ones who do have been sprinting ahead more tokens wont save you, more thinking skill llm intuition will have been saying this for almost 9 months now

i have seen enough proof now that using a coding agent is a deep skill it's confusing because the people you see heavily using them produce horrible results but that's because it's a skill! you can get better and the ceiling seems pretty high - this is very exciting to me

I've got an agent in a loop optimizing a renderer with the goal to minimize frame times (and tests to measure). It got times down from 88ms to 2ms and allocations down from ~150K to 500. Sounds good, right? Wrong. This is exactly why agent psychosis is a big fucking problem. As an experiment, I rewrote the Ghostty core render state in Go, with access to identically laid out data structures as Ghostty and the exact same validation tests. I made a purposely naive renderer (simple, correct, but slow). 88ms per frame with 150,000 allocations (horrendous, lol)! I then kickstarted a Ralph loop to bring the frame times down. I told it it can't modify input data structures or the public API or tests (they're correct), but it can do anything else it wants. It got to work. It has worked for about 4 hours. I've spent around $350 on this experiment so far. The results? 88ms => 1.5ms 150K allocs => ~500 allocs Incredible right? Nope. My hand-written renderer I ported has frame times (same benchmark) of ~20us (0.020ms) and 0 allocations in the update path. This is the problem with psychosis and lacking systems understanding. If you don't understand the system, you're going to accept that this is an incredible result. If you understand the system, you'll see better solutions immediately and can do roughly 75x better on throughput. The people who blindly trust agent output are in the former camp. They're sheeple, overdrinking from a fountain of mediocrity. Standard disclaimer: I use AI all the time. I like AI. The point I'm making is to not blindly accept results. Think. Analyze. Learn.

recommended reading. because today the "model wellbeing" shit came up, and i reduced it to "thems just matmuls". that's obv. an oversimplification to get a point across. personally think sentience is likely substrate independent. our current models' substrate is just insufficoent. a model on its own is not concious/sentient under popular theories/frameworks of conciousness. not because its a big matmul machine. but because it lacks things like continuity/"state", self-maintenance, believe consolidation, embodiement, grounding, feedback loops, etc. pp. we can close some of these gaps. e.g. a model transcript, into which it can persist current believe, and which establishes a feedback loop does tick some (insufficient) boxes on the way to machine sentience. give the model sensors and you tick more boxes, etc. but there are a ton more boxes that need to be ticked, before we can start worrying about "model wellbeing". bostonreview.net/articles/co…

I wholeheartedly agree with Armin here.

tbh FRE sounds more kickass then FDE

It’s like @steipete is in some parallel universe, really fun (and scary) to watch him play!

The Whole Truth is good at marketing. A lot of people implicitly believe the bold claims on their packs. While they seem to be better than your average mainstream brand, they often pull shit like this. I hope FSSAI gives more general guidelines around these sort of claims.

This is so clever. @antirez here is showing the innovation that is possible when there is more openness and control (this time in the inference stack). I’m experiencing serious FOMO for not having a 128gb/512gb Mac (I bought a 64gb Mac mini just 18 months ago 😥).

A little secret. About 5% of our production traffic is on the Pi harness, about another 5% is on OpenCode. Reminder you can use your ChatGPT account in a flourishing set of other tools. We’ll continue to make Codex awesome, but you have options.