Cole Brown

@VictorTaelin Do you have projects you wish existed in the Bend ecosystem? I’ve been an admirer since your early HVM/Kind days. I have a dependently typed lang bg too. Could prob find some budget for experiments like this.

I just realized that, for the first time ever, progress on both HVM and Bend is bottlenecked not anymore on time, and not on my own knowledge, but on money. I could, right now, be throwing thousands of agents on the repos, with well crafted prompts, and progress would accelerate 10-fold overnight. I know "my agent is outputting 100k lines of code" is a silly meme, but it does apply here. Bend is now equipped with everything it needs to convert agent time into real progress, at any scale. This weekend I validated that, leaving agents working on the standard libraries overnight, and I woke up to thousands of lines of well written, tested, verified code that just worked. One of the agents even started implementing a 3D rendering library in Bend, and it got incredibly far! So, if I can trade money for having ThreeJS in Bend's ecosystem overnight, I think it is not a stretch to say we could port nearly anything to its libraries very quickly. Now, of course, our budget is limited; currently, this is all running mostly with a few agents under a single personal Codex/Claude subscription. But what is preventing me from just throwing, say, 10k on it, and checking what they create, and how far the go? Nothing, and I think this is the next step. It is worth noting that progress here isn't in the "solving new problems" sense; they are not creating anything "new"; they are just implementing existing things in a new language. But that's exactly the "Bend ecosystem is small" problem I'm trying to solve right now! Seems like AI was made for this

Thought this was a @tbpn graphic

@VictorTaelin I am completely sure this is Gemini lol. It’s always telling me I’ve uncovered the holy grail of ___.

This is a fantastic and deep observation. You have correctly identified a fundamental knot in dependently typed language design. Guess the model!

@emollick Well I replied to the old tweet not the new one, but maybe it’s just good proof that I’m human 😂

@emollick With MiniMax M2.7 appearing to be proprietary, this looks even more accurate. I suppose there is now the big question of: does this make it easier for these labs to reach the frontier? I’m sure many compromises on eg param size were made for open models. But is that it?

I think it is entirely possible that there will be no new frontier open weights models at some point in the near future. Counting on the Chinese AI labs to keep making their models free forever doesn’t make sense as model costs rise & the value of having a frontier model goes up

@mattyglesias This is something I’m shocked the left is still missing. Over regulation slows progress. Markets and market incentives can work wonderfully.

Solar progress depends on sunshine, transmission infrastructure, and the ability to obtain large parcels of land — it's not about anti-fossil fuel regulations.

@Teknium Ok cool! I idle there so I’ll speak up :)

@dtcb I dont use honcho myself, but happy to take a look, Eri in our discord is a great representative and dev from platisc labs who builds Honcho who could probably help though!

@Teknium Yo! Been enjoying using Hermes almost as a replacement for a shell for my devbox. It's been rad. I've made a small hack to support self-hosting Honcho, and I've got it running locally. Do you deploy it yourself? I'm noticing some really odd context bugs...

@Teknium Basically sometimes it seems like Honcho is injecting context that focuses on something way back in our session and completely ignores the latest instruction. Would be awesome to have more visibility on what Honcho is doing. Maybe I'll hack that together 😂

@scaling01 Guessing there is some GTC timing? Interesting to see small jump up to 119B parameters. Very sparse, too. Definitely learning from Qwen and Nemotron.

but why not post on main Mistral acc?

Mistral 4 has been live for 30+ minutes but I haven't seen a single post?? western open-source is really cooked huggingface.co/mistralai/Mist…

@patdennis These same folks will, in the same breath, say it’s too powerful and must be regulated out of existence, lest it destroy white collar work forever.

I have not been stunted by offloading cognition onto excel. Keeping cells up to date manually takes a lot of cognition and is rather focus on the big picture

@swyx @natolambert It’s sad, especially with Qwen3.5 being, in my opinion, an incredibly under hyped release. Really hope the researchers land somewhere they can continue their work.

i am actually still not over how Qwen as we knew it, one of the S tier Tigers, is over.

@badcryptobitch Gna give ‘em a go tomorrow!

@dtcb Haruomi Hosono's SFX Yukihiro Takahashi's What Me Worry

YMO

@badcryptobitch Ooo got a rec?

@dtcb Fellow YMO enjoyer!!! Don't sleep on the other members' solo discography either

@badcryptobitch If you haven’t dipped in, “The Thousand Knives Of” is mind blowing!

@badcryptobitch All Ryuichi Sakamoto work tbh

One unexpected ER visit is all it takes to remind you of the US’s shortcomings.

@swyx Is this not just another instance of buying distribution? In that case, probably better to price per monthly download or something? Or are you more seeing that OSS really derisks the diligence stage of hiring someone competent to solve a problem @ the labs?

btw if you can build a category leader open source project in ai engineering right now the market acquihire rate is ~$10-$100m per ai engineer. you do not need to figure out a business model, you do not need GTM, you do not need funding. just build things clankers want.

@PrimeIntellect Any chance these sessions will be recorded? Just had sinus surgery and can't fly 😬

Prime Intellect Day Saturday — March 14 (Pi Day) 4-11pm — The Melody of SF Talks on frontier model training, agentic RL, computer use, coding agents, diffusion LLMs, automating AI & science, robotics — and why the singularity will not be singular.

@swyx @sama For sure. The advantage y’all have is palpable. Devin is such a complete product already, whereas other cloud agents have some ground to cover. Just wanted to share—it’s a real joy to use.

"Build a company that benefits from the models getting better and better" — @sama devin brain uses a couple dozen modelgroups and extensively evals every model for inclusion in the harness, doing a complete rewrite every few months. hearing a lot of "devin is good now" feedback but its largely the same process that the team has been running since @ScottWu46 bet on cloud agents in November 2023. agents are really, really working now and you had to have scaled harness eng + GTM to prep for this moment