ttul

That team was mostly hired away by NVIDIA.

I built this in 2001 (https://pyinline.sourceforge.net/). Obviously, that ancient thing doesn't work anymore, but the idea worked then and it ought to work today. That being said, I think there are significantly superior options in existence now!

That is correct. The gpt-6 models are strange.

We're so far beyond lines of code improvements. Agents are building entire features with mergeable PRs that we now deploy with confidence. That was impossible six months ago and unimaginable a year ago.

So, my measure is not just line of code. It's deployable features. And yes, by that measure, we have gone from zero to "lots" in a very short time.

I don't see why a payoff would have appeared yet. Many (but certainly not all) software companies have increased their output without hiring more people. But having software with more features doesn't necessarily drive immediate productivity gains in the rest of the economy. It will take time for those benefits to percolate. We've only had access to models good enough to majorly accelerate code writing since early-2026. I feel like we're in the very beginnings of the first part of this wave.

For sake of comparison, consider e-commerce. As a proxy for "impact on the economy", we can use Amazon's revenue. They went public in 1997 and that year had revenue of $148M (just 0.14% of Walmart's revenue that year). By 2000, revenue was $2.76B (1.65% of Walmart). Bricks and mortar didn't start feeling anything until probably 2011, when Amazon crossed 11% of Walmart's revenue. And it wasn't until 2025 that Amazon finally surpassed Walmart.

But, if you worked in tech in the e-commerce business, things were already feeling pretty hot by 2007. And at that point in time, the e-commerce sector was still miniscule compared to traditional retail (~4%). Software is very exciting these days and the most fantastic things are happening. They will eventually percolate out to the rest of the economy. But it will take many years before everyone notices the difference.

These full-duplex models are getting very "Her"-ish lately. gpt-live-1 is fantastic. Dessa, though, seems to be aiming at full human parity. There's a huge market for AI girlfriends that they should be looking into in earnest...

From the site:

Q: Do the iServices like iCloud, App Store, etc work? A: No, at least not yet. They possibly require something iOS refers to as the “Legacy FairPlay Key Descrambler” AES (hardware) command.

Extremely subtle hint. Well played.

Pack it up and go home, everyone. The alpha-release doesn't chart. Time to throw in the towel. /s

Grok can be fast because X.ai does not worry about gross margins yet. They can tuck those nasty secrets away deep inside SpaceX…

I suspect that Dots and Duet are very high margin drivers for the labs. Personal assistant chat sessions don’t have to be realtime, so workloads can be scheduled on low priority batches and be waited if higher value work comes into a queue.

Inference providers are always trying to fill inference batches and having a stream of low priority requests can help pack batches that might have gone empty, without raising the latency of higher priority requests such as live chat sessions.

While Dot and Duet certainly need to be useful to work out as products, I think the gross margin mechanics are just as interesting.

I find it is okay if I tell the agent that running the formatting step is required before every commit.

I've noticed the line breaks issue. It's easily fixed with a formatter like prettier and, frankly, that kind of quality-of-life thing is best handled deterministically anyway.

Surely they mean Opus-5.5...

Super cool. I'll have to give it a whirl! Keep up the amazing work.

I am inclined to think this way in a commercial setting, but for open source contributors, where their work was done for the love of coding, I can understand feeling like the end has arrived. In my company, I'm building the mechanics of factory-coding and trying to get the team think in higher-level terms: what needs to be built, what does it need to do when it deploys, and how can we build the factory that gets that done repeatedly? To me, building the factory is an interesting problem and agentic coding does not eliminate the higher-level design that absolutely must be done, all of which is, to me, the interesting part of software dev.

But none of this is going to bring back actually writing code for the love of it. That will become a hobby kind of like knitting or sewing.

I'm curious if you could comment on why your DeepSWE performance is below Haiku. This is not a dig - getting 46% vs. Haiku's 52% is still amazing. I'm just wondering if perhaps your post-training was less aggressively RL'ed on coding tasks than perhaps Haiku was. Having a cheap 1,000 tok/s option for coding work is a dream because - if it's reliable enough - you could blast through work such as high-level security reviews or finicky and repetitive refactorings that don't necessarily require SOTA performance.

I was chatting with a friend who commits on a very well known open source project (~80K stars). He is disillusioned not because the robots aren't good at what they do -- they totally are. The sad thing is that he enjoyed writing code and working with the community on improvements because it was a very human, collaborative thing to do. So, while the robots are great at pushing the project forward even faster and doing fantastic things for security and performance, the joy is gone.

I can totally identify with that feeling. It must be how professional mathematicians are feeling as well.

It's a little more involved than just having proxies in Pakistan. I'm quite sure Google has some protections against bots and scraping and SerpAPI's team unfailingly seems to work around those restrictions.

Agentic coding lets you build more things in less time. Companies have responded by ordering product teams to pick up the pace of releases. But the bottleneck isn't coding execution, it's choosing what to build and verifying that what got built doesn't suck.

I think this problem will be solved. Frontier models recently learned how to do 3D really well (Astra was first on 3D; Opus-5.5 nailed it weeks later). How long before they get trained up on product management? It won't be long. Sorry.

I'm serious. Linus does not rule Linux because of some cabal of dark overlords. He leads because he continues to deserve the job. I think Linus would be the first to admit it's time to retire if he started failing in his leadership of Linux development. That being said, there is truly nothing standing in the way of anyone forking the kernel and building their own version, over which they have complete control. So, I think it's an entirely fair suggestion to make.

SerpAPI is comically good at evading Google's scraping protections. If what you want to know is, precisely, what someone searching Google in Pakistan would see, this is the API for you.

My read on Terence Tao's comments of late is that professional mathematicians are principally upset about AI-submitted math papers lacking the kind of broad teaching content that high quality human math papers offer. While the answer might be correct, AI math papers don't teach the reader in the way a human author would have. Professional mathematics is not just about solving problems; it's just as much about spreading knowledge so that others can program their human brains with the same understanding that the paper's author had when they wrote the paper (and achieved the result).

In other words, the crisis in mathematics (as I read it) is that AI can solve without teaching. The result may be impressive, but if the only beings that can move math forward in future are machines grinding away on billions of scenarios that no human would have the patience to undertake, this leaves those in the profession wondering about their place in it.

I suppose, feel free to fork the kernel and take the lead?

At least one betting market has the odds at around 25% [1] of a direct military conflict with a NATO country by year's end. It will be interesting to see how things change after November's midterm elections. For what it's worth, Metaculus forecasters show a spike of likelihood at the end of September, 2027, with about 20% of forecasters saying that's when a military clash will happen. I'm not sure what's driving that spike in the forecast.

[1] https://polymarket.com/event/nato-x-russia-military-clash-in... [2] https://www.metaculus.com/questions/41238/when-will-russia-a...

2 billion per day is 23,148 per second on average. Perhaps peak bursts are >100K per second. That's a seriously insane number to keep up with at GitHub's scale. I've worked on systems that write >100K things per second and when it has to be globally distributed and relatively consistent within seconds, that's a very tough challenge.

TOTP is susceptible to man in the middle attacks. Passkeys are not. That alone is a solid reason to favor Passkeys!

I should frame this for the wall behind me when I run meetings with my team. It’s the core problem with AI in coding.

Dan Kaminsky told me about the signing ceremony years ago (bless his heart, he was one of the people trusted with DNS security). Apparently, everyone on the signing committee flies to a central location carrying a hardware key. Then they take turns inserting their hardware key into a hardware security module. It's all done face-to-face because that's the only way to guarantee that human consent was granted properly at each step.

Each ceremony is recorded on video and distributed to the public. The last one was in August, 2026: https://www.youtube.com/watch?v=-QqYS3oLfL8&t=1s

Thanks for clarifying. Not sending a warning for a $34K bill is pretty terrible. My company runs a service that allows larger customers to run overages and all of our bad customer experiences have stemmed from surprise overages that we failed to warn them about. Over the years, we have learned to be less stupid about that!

It appears that even very large companies like Cloudflare may need to brush up on customer experience from time to time.

A $34K bill with Durable Objects implies they somehow made 227 billion requests in a month. Sure, if that's the kind of traffic you're serving, your bill ought to be $34K! That's 87,000 per second!