Shoggoth Jockey Suma Corn Loude @ Dunning Krugers University - also farmer
Magic Kingdom
Joined February 2011
- Tweets8K
- Following1.8K
- Followers495
- Likes9.4K
looks like our embryonic research team landed 2nd in the first leg. I really enjoyed working on this in a more collaborative manner! If I am at a casino, I am playing craps, because I want to high five, and win together. Really happy the game here is embryo genesis extrapolation
Frosty40 retweeted
Some early thoughts on Claude Duet:
- Claude Duet acts as its own independent app, not a part of Claude. They are connected but independent.
- Which also means you literally have to download a different app.
- This is a huge L for me personally because it limits its utility
- You cannot access Duet on Desktop App 🫠
-You must goto browser to use it on your computer. You can't see tasks, nor can you hand off tasks with Claude, nor is it aware of Duet whatsoever.
- Duet can read your cloud conversations and your Claude Code sessions, provided you have Remote Control on, but it cannot interact with those threads. It can start its own tasks based on them.
- Because it runs on Opus 5.5, its unsurprisingly really capable.
- It creates tasks lists which you can track, which are pretty nice. It also has "suggested tasks" which can be helpful.
- The plugin ecosystem is rich, and includes Link/Shop Pay.
- Plugins are shared, which is nice so you don't have to reconnect everything
- You can create different conversations that are scoped (similar to Grok Bots)
- Duet is quite proactive and will surface tasks and remind you of deadlines, check emails, etc.
Verdict: right now, I genuinely feel this is the least useful personal agent, although it can be remedied with better integration into Claude itself.
Even Dots is significantly more useful, and people meme on it all the time!
Dots I can access from every device: web, mobile app, desktop app and all the conversations are synced.
> Dots can create tasks on any device.
> Dots can work in shared spaces
> Dots can be tagged directly in supported apps
> No separate apps
I actually think Dots are way better than people give them credit for, but I think most of Dots' utility is realized when working with a large Team in Slack.
Overall, for me Duet is a complete miss, and you shouldn't feel FOMO to get in. I am sure it will improve, but I don't like how compartmentalized it currently feels.
Hopefully once out of research preview, they will integrate it more deeply with the Claude apps.
Frosty40 retweeted
The computers we depend on should be computers we build ourselves.
BlueChip. Built in America.
Frosty40 retweeted
Uploaded 693 llm-inference-bench throughput benchmark results from 228 distinct models onto LocalMaxxing
The goal here is for a unified source of LLM throughput benchmarks recorded across a matrix of concurrency levels and context lengths, supporting a wide range of hardware, while allowing community contributions
These were validated on NVIDIA H200 and RTX PRO 6000 Server Edition nodes, using either 1x, 2x, 4x, or 8x GPU configs, and across both SGLang & vLLM
Benchmarks were taken from older & newer models within the following 35 orgs:
Allen AI · Arcee AI · Cohere · Datalab · DeepSeek · Dots Studio · Google · IBM Granite · Inclusion AI · Inco AI · Inferact · Intel · Liquid AI · Meta · Microsoft · MiniMax · Mistral AI · Moonshot AI · Nanbeige · Nex-AGI · NVIDIA · OpenAI gpt-oss · Paradigma · Prime Intellect · Poolside · Qwen · RadixArk · Red Hat AI · StepFun · Tencent · Thinking Machines Lab · Xiaomi · Z Lab · Zai · Zyphra
A huge thanks to the incredible crew over at @YourLocalAILab for creating this benchmarking tool, @LottoLabs for adding this page to LocalMaxxing, and @verdacloud for the compute
You can find these and other community uploaded runs at:
localmaxxing.com/en/inferenc…
Localmaxxing.com now has llm-inference-bench integration and matrix style run support
Lets you see how models act under different loads and concurrencies on one page
OG cards and shareable links asap
Download localmaxxing-cli and try it out
I’m going to be populating as many models as I can with 1xP150 and 2x3090
Alot of you are just converting your entire codebase to rust. have fun, your doing it right, perfect timing, no reason to think that's not a good idea. 6 changes 10 minutes per compile no big deal, just an hour, its fiiiiine
Frosty40 retweeted
Last week I said I still want models open even though anyone can strip the safety training off.
This weekend, running local AI on my Mac mini, 11 open-weight models rewrote the Gettysburg Address 19,019 times and the two biggest kept finding their way back to Lincoln.
What I learned:
1) GLM-5.3 Flash from @Zai_org and DeepSeek V4.1 Flash from @deepseek_ai looked rock stable until I read the outputs. 86 of GLM's 1,000 rewrites open with "Four score", and by round 1,000 both were closer to Lincoln's original wording than at round 1. Part of what looked like stability was memory.
2) The small models had nothing to fall back on and drifted off. DeepSeek-V2-Lite turned the 278-word speech into 13 words about advancing your career, and Qwen3.5 0.8B ended up on project planning. I think it's the same pull the big models felt (each one slides toward whatever its training data made most familiar), it just landed on workplace topics instead of Lincoln.
3) A single score would have hidden all of this. I only caught it because I could open the files and read them.
4) Running locally meant the weights never changed mid-run and I paid nothing per rewrite. Anyone with a Mac can rerun it and check my numbers.
5) The two biggest models don't fit in 64GB, so they streamed off my SSD and ate about 36 of the 45 hours. They still ran. Most of the others finished 1,000 rewrites in under an hour. The smallest did 350 tok/s in under 1GB of memory, and Qwen3-Next 80B, an 80B mixture-of-experts, did 73 tok/s in 45GB.
6) Using a famous text baked memorization into the result. The next run adds a text no model could have seen, so I can separate stability from recall.
The very reason why I want weights open. A benchmark score wouldn't have shown this and no API would let me run 19,019 rewrites for free. All of it is public, so you can check my work or find what I missed.
If you've run long chained tests on your own machine, I'd love to hear what you found.
Anyone with a GPU can strip the safety training off an open-weight model for a few hundred dollars and nobody can take it back!
I still want these models open. 9 days of running them on my own Mac mini made me more sure of that.
A few things I believe now that I didn’t a year ago:
1) Local AI will carry a lot more inference than people expect. The gap between my desk and the cloud is still there, but it shrinks with every good open-weight release.
2) Qwen3 30B from @Alibaba_Qwen does work on my 64GB Mac mini that I would have paid an API for last year. I care more about owning a model than about its benchmark scores. I can examine it, adjust it and keep using it after the company that made it changes its mind.
3) Most of what I’ve learned came from builders sharing their setups. The 24 people tagged below (and everyone on my list) taught me more than any research paper 🙏
4) Qwen3 30B runs faster on my Mac than Qwen3 8B, 106 tokens per second against 61, because it only uses about 3 billion of its parameters per token. I didn’t expect that, and @PrismML’s Ternary Bonsai 2 27B surprised me even more. It stores almost every weight as -1, 0 or 1, fits in 8.4GB and still runs faster than Gemma 3 27B from @googlegemma, 26 tokens per second to 18.
5) Openness is still our best bet against people who misuse it. Security improved because defenders could read the same code attackers were reading, and I’d rather AI work that way than leave the strongest tools with whoever controls the API.
If you run models on your own machine, I’d love to know what setup you use.
Frosty40 retweeted
GPT-6.1 Terra means business.
Our new fast, efficient model pairs sharper visual reasoning and precise computer use with clearer, more natural writing.
From decoding technical drawings to drafting proposals, filling forms to editing videos—it’s built for the whole workday.
GPT-6.1 Terra is rolling out today to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.
🤖 Made with AI
Frosty40 retweeted
A lot of awesome progress has been made in last one week on an array of frontier scientific problems @yukonresearch, ranging from quantum computing to cryptography to biology.
I want to use this thread to summarize the progress.
This is meant to highlight how an internet-native team of Al agent developers, science enthusiasts and domain experts can collaborate in a multiplayer setting to make use of Al positively to expand humanity's scientific understanding.
1/n
Frosty40 retweeted
Introducing Fern, Mirendil’s first model: pretrained from scratch, with up to 10M tokens of context!
That’s ~0.5M lines of code (roughly Megatron-LM scale) or days of an agent’s work, all in context at once.
Its first version, Fern 0.1, had a short round of post-training 🚢
AngelX will be going hard in the paint with muse, full Muse mastery going down, pushing FB models to the limit with what it can do.
