←
mteam.gwei16,854 posts
mteam.gwei avatar

mteam.gwei

@mteamisloading

cofounder @spire_labs ~ eliminate the middleman

◷ Joined November 2021▧ 853 Photos and videos
mteam.gwei avatar
mteam.gwei@mteamisloading·29 Oct 2024

I'm in crypto to eliminate the middleman. Disintermediate all human coordination. Finance, commerce, social, governance (corporation, even state level), information curation, internet infra, physical infra, content creation, science, property rights, employment...

Daniel Khashabi 🕊️ avatar
⟳ mteam.gwei retweeted
Daniel Khashabi 🕊️@DanielKhashabi·Nov 17

AI agents do remarkably well in discovering new ideas and solving hard science problems. Do we understand how LLMs search their way to a final idea? How do they recover after getting stuck during ideation? We study LLM ideation by turning it into a two-agent game of question-answering with binary feedback.

Jackson Dahl avatar
⟳ mteam.gwei retweeted
Jackson Dahl@jacksondahl·Nov 17

My current feeling is that frontier OAI/Ant/Gemini models with harness on desktop are phenomenally useful for real work... But despite how useful they are, I also still feel like my mind is being psychologically raided when I'm switching between all the different chat threads And on top of this, all of these agents I'm supposed to text all day that seemed awesome at first are in fact not very useful (dot/muse/instinct/whatever) Makes me increasingly convinced that "chat" alone is decidedly not the final interface of using a computer, or at least doing work on a computer. So much power and so little interface innovation Curious for those who disagree. Perhaps I'm just bad at this

Theo - t3.gg avatar
⟳ mteam.gwei retweeted
Theo - t3.gg@theo·20h

Thought more about this. Anthropic’s lead started with Fable 5. It was the first true “next gen” model. I honestly think most of yall missed just how far ahead Fable was at launch. Understandable tbh. It required you to use the model entirely differently. It was prohibitively expensive. It got taken from us after just 3 days of usage. It had no ZDR policy, so you couldn’t even use it at work in most places 🙃 The result: most people perceived the gap between Fable 5 and GPT-5.6 Sol to be small. IMO, it was massive. GPT-6 Astra mostly caught up on raw capability for short term tasks, but it was unable to stay coherent for long runs. Meanwhile, Opus and Fable keep getting easier to work with and more coherent on long runs, while also gaining in capability. Opus 5.5 brought the Fable wins to everyone, by making it way cheaper, more generous on sub limits, and also great for enterprise (zdr!) OpenAI has maintained pace with capability, at the cost of coherency and readability. Tl;dr - Anthropic has been ahead since Fable 5. OpenAI’s attempts to catch up on benches have hurt the model pretty much everywhere else

It’s hard to put into words just how far ahead Anthropic is right now

mteam.gwei avatar
mteam.gwei@mteamisloading·Nov 17

get your agents to do unifying attention, this is the frontier

today, I invite you to notice that your habitual way of paying attention is one of analysis, which separates and discriminates between objects there’s another kind of attention that integrates and unifies whatever you attend to what’s it like to drop into unifying attention?

mteam.gwei avatar
mteam.gwei@mteamisloading·Nov 17

whats the best place to get low latency high throughput LLM tokens for cheap? talking 200+ tps

ThePrimeagen avatar
⟳ mteam.gwei retweeted
ThePrimeagen@ThePrimeagen·Nov 14

If you ever wonder why my only goal is the enjoyment of technology is because I have been forgiven. I have nothing to prove or to win. No X arguments or tech arguments to assert why I am the best, nope. I simply get to enjoy software and live the purpose my Creator gave me. I believe my "spiritual gift" is creation. I love creating and want to share it with everyone. Jesus saved my after 19 years of decently destructive behavior. I would trade everything for this gift.

Jesus saves and God is good. 🕊️ I don't talk about it much, my faith is a big reason I feel so strongly about people who scam, cheat and steal. Right now there's a culture of people who think it's the only way to get ahead in our world. It's not only misguided, but self-defeating. It works short-term but it always ends the same way: My son, if sinful men entice you, do not give in to them. If they say, “Come along with us; let’s lie in wait for innocent blood, let’s ambush some harmless soul; let’s swallow them alive, like the grave, and whole, like those who go down to the pit; we will get all sorts of valuable things and fill our houses with plunder; cast lots with us; we will all share the loot” my son, do not go along with them, do not set foot on their paths; for their feet rush into evil, they are swift to shed blood... These men lie in wait for their own blood; they ambush only themselves! Such are the paths of all who go after ill-gotten gain; it takes away the life of those who get it. Proverbs 1:10-19

will depue avatar
⟳ mteam.gwei retweeted
will depue@willdepue·Nov 12

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

Post attachment
Richard Davey avatar
⟳ mteam.gwei retweeted
Richard Davey@photonstorm·Nov 11

I asked Fable 5.2 to "go wild" making a game, and all I got was this Pong clone. And then I played it ... 👇

Post attachment
Simon Willison avatar
⟳ mteam.gwei retweeted
Simon Willison@simonw·Nov 10

There is a mysterious intermittent beep coming from somewhere in your room of unopened boxes. Do you: 1. Open boxes methodically until you find it 2. Vibe code software to triangulate it using two laptops, two iPads and a phone all listening at once simonwillison.net/2026/Nov/9…

mteam.gwei avatar
mteam.gwei@mteamisloading·Nov 10

Say hi to Yosef, a personal context system with a simple idea: ingest as much information as possible and dream it into dynamic memory pages that load into agents based on their prompt. A replacement for manual skills & traditional memory. Inspired by @victortaelin's UniiChat.

1:12
DANN© avatar
⟳ mteam.gwei retweeted
DANN©@DannPetty·Nov 6

Vectorize your stuff with Quiver, y'all.

0:45
Howaboua avatar
⟳ mteam.gwei retweeted
Howaboua@Howaboua·Nov 6

I am beginning to doubt that gpt models can actually understand what subagents are. 1. They keep issuing parts of what harness exposed as a part of the prompt. Dude, the subagent has the same capability to interact with its tools, it doesn't have to be told how an already exposed toolkit works. 2. They keep telling subagents about stuff that is completely unrelated to their task, but has happened somewhere during the main session. My next step is trying to figure out how the hell to minimise this. I can't remember if I'm passing these prompts as dev role or user role, but either way, this kind of overprompting most definitely leads to poisoning a clean agent with BS instructions. But what does one do, put in "other agents arent idiots, they do not need to be informed how their tools work?"

Georgios Konstantopoulos avatar
⟳ mteam.gwei retweeted
Georgios Konstantopoulos@gakonst·Nov 2

you should never let a cache distract you from fixing the underlying. in distributed systems you want to use a cache to hide the latency between components. but you still want very fast reads w/ writes passthrough in the bg to make the cache management easier.