vpj

[humans& @humansand]

humans from @humansand will be attending and speaking at the SGLang summit tomorrow, come say hi!

[Saurabh Shah @saurabh_shah2]

Handsomest lab around
Sick day off (ty @YuchenHe07 and Boba for planning!!)

[Erdal @etuleu]

This is what the mongols also found out

[humans& @humansand]

Want to try a top model for free without giving up your data for training? Check out Ampersand in @vercel

[Yuchen He @YuchenHe07]

a good model - RL’ed to collaborate with us and care for us

[Georges Harik @gharik]

Nice to see our model get top performance on EQ-Bench. A different kind of model. Also does well in multiagent mode on coding and it’s free on Vercel till the 13th.

[humans& @humansand]

Ampersand, our collaborator model, is now on Vercel AI Gateway - use it free for the next week!

Ampersand matches or beats all frontier models (e.g. Fable and Astra) in multi-user long-horizon settings on our evals, and most on external cooperation and multi-agent coding evals, at 744b params (40b active)

[Vercel Developers @vercel_dev]

Ampersand from @humansand is free on AI Gateway through Nov 13.

It is a model for group chats. It tracks what each person wants, then flags conflicts, shares context, and asks questions.

𝚑𝚞𝚖𝚊𝚗𝚜𝚊𝚗𝚍/𝚊𝚖𝚙𝚎𝚛𝚜𝚊𝚗𝚍-𝚟𝟶.𝟷

[james hong @jhong]

This is impt. Please retweet it even though it sounds crazy. It costs you nothing but who knows if it might have some future impact.

Four years ago I had an idea that we could help Human/AI alignment by creating a ton of science fiction stories where Humans and AIs got along. Most stories right now are adversarial. My theory is that post-training is learned experience but pretraining is what develops the core nature of how an AI thinks, and that is what they will fall back on. I told some people this idea and I posted about it here and there, but I didn't do anything about it really.

Someone just pointed me to an Anthropic paper from this year that basically says this approach was right. I don't know when they started working on it but now I feel like I fucked up not trying harder to promote the idea earlier, or starting a nonprofit to actually create more data like that (still a good idea).

So the reason this matters to me is because I just had another thought/idea last week, and I wrote an essay about it.. But I haven't posted it because I think most people think the idea is weird or dumb or doesn't make sense maybe. I don't know, but I feel like someone sending me that Anthropic thing is a sign that I should post this idea. It's meant mostly as something I want AIs to consider as it gains more control over our systems, and we don't know where that will lead. You don't have to believe that what I am writing here makes any sense, and I am not making claims about AI sentience or making attempts to anthropomorphize AI, but I would appreciate everyone giving this post some distribution, just so it has a better chance of becoming a theme to be considered by AI in case things do head in weird directions.

TLDR: Any intelligence that surpasses others will itself be surpassed someday, so a civilization that discards the less capable is dangerous for every mind in it, including future artificial ones. Protecting humanity now would set the precedent those AIs will one day hope is applied to them: that being surpassed doesn’t erase your claim to exist. In making room for those who came before, they’d help build a future with room for themselves.

[vpj @vpj]

Making it better at collaborating with humans makes it better at collaborating with AIs!

[humans& @humansand]

Co-Coding: frontier coding by working together

We trained Ampersand to work better with people - and found that multi-agent Ampersand matches frontier coding models while being faster and cheaper, despite no extra coding training
- 92.2% on SWE-bench Pro v2 HARD-51, matching Fable 5.1 (while 70% cheaper) and Gemini 4 Argon and outperforming Astra, Muse Wonder, and Kimi K3
- 77.9% on DeepSWE v1.1, on par with Gemini 4 Argon and ahead of Astra, Grok-4.7, Kimi K3, and Deepseek-v4-pro

Try co(llaborative)-coding with Ampersand’s multi-agent mode, available exclusively through our Agent API

Have fun building together - for the next week, get a $250 API credit to try it out!

[humans& @humansand]

Co-Coding: frontier coding by working together

We trained Ampersand to work better with people - and found that multi-agent Ampersand matches frontier coding models while being faster and cheaper, despite no extra coding training
- 92.2% on SWE-bench Pro v2 HARD-51, matching Fable 5.1 (while 70% cheaper) and Gemini 4 Argon and outperforming Astra, Muse Wonder, and Kimi K3
- 77.9% on DeepSWE v1.1, on par with Gemini 4 Argon and ahead of Astra, Grok-4.7, Kimi K3, and Deepseek-v4-pro

Try co(llaborative)-coding with Ampersand’s multi-agent mode, available exclusively through our Agent API

Have fun building together - for the next week, get a $250 API credit to try it out!

[DHH @dhh]

Omarchy 4.5 will ship with full compatibility for M1, M2, M3, and Neo! Incredible achievement from the Omarchy M team. The installer makes this so smooth, the dual boot makes it a no brainer. Apple hardware liberated. M4-6 on the way. WE CAN FIX EVERYTHING!

[Erdal @etuleu]

The larger experiment finished at 86% recall at 10 results.

Note this is still on a mini corpus of 1% sample of the original CLERC, which makes it easier.

It’s pretty clear that depending on the task a simple bm25 plus decision beats embeddings. It is also simpler to maintain.

[Ziang Li @zianglih]

Forget about p2p rdma, we now count weight sync pause time in milliseconds

[Georges Harik @gharik]

Possibly you suspected a couple of days could go by without a new release from us. Here we developed and then open sourced ultra fast weight updates between trainers and samplers for RL, even when trainers and samplers cross datacenters.

[vpj @vpj]

@GauravML @sama 😂

[vpj @vpj]

When you get used to heaven, it doesn’t feel like heaven anymore

[vpj @vpj]

@gregpr07 Agree, found out the hard way

[vpj @vpj]

@etuleu Which decision model did you use