@NathanZ721 This is really impressive! Love seeing specialized models crush it on niche tasks like this. Is the model publicly available yet, or any chance to try it out somewhere?
English
kenw
89 posts
I trained an SVLM to do one thing really well: read music
Sounds simple, but frontier models consume millions of tokens over hours to write scripts and analyze glyphs, only to hit <50% accuracy.
My model takes 10 seconds on one GPU to hit 98% accuracy
English
We found compacting agent context on cache misses reduces token costs by over 50%.
The graphic is a real Fable 5 trace from our CTO, where we simulate cost would before/after compacting after periods of over 5 mins of inactivity.
By keeping context within 128k-256k tokens, spend goes from being quadratic to linear in the number of turns.
The only problem is self-summary takes 1-2 minutes on big traces! We trained an absurdly fast compact model to get the cost savings without harming the user experience.
Relace Compact is a modified LLM designed for token-level classification, which means we can run it at 50k tok/s and an order of magnitude lower than cache read cost ($0.20/million input and $0.20/million output tokens).
Try it out in Jacq, or visit our docs to integrate into your own product!
English
@bitbangit @nateparrott Hi! We just added png / jpeg export options for images (you can access under share menu with more formats, or the bottom of the edit panel). Let me know if those work for you or you are looking for something else.
English
folks who’ve tried Claude Design — how was the experience? what works well and what could be better?
English
@Daniel_Farinax I think this is actually a good idea. Make it similar to API batch mode; no guarantee of a quick response for 50% of the cost. Would be great for long running tasks that aren't time sensitive.
English
Live tweeting video editing of my AIE talk.
Swyx sent me 4 videos that are about 60GB, I also threw in my HTML deck.
This is my starting prompt.
Thariq@trq212
got the raw footage of this talk from @swyx - 59 gb going to see how Fable does editing it
English
@blind_via It's obvious you've never lived in the hood. You might have ppl going door to door trying to sell you steaks/tide/perfume/etc. they lifted a few hours ago from your local supermarket.
English
Who buys soap and toilet paper from the guy down the street? Fencing stolen pens and toiletries sounds crazy
sarahsecrets - Zeppleinætrix@sarahsecrets6
@blind_via shoplifters steal items that they can turn over quickly, but are not worth enough to warrant major police investigation. first it was razor blade refills, then soap and toiletries now it is stationary welcome to America
English
26 sec… damp sponge got in the way.
i2cjak@i2cjak
I AM PROUD TO ANNOUNCE ROUND 2 OF THE OFFICIAL AMERICAN EMBEDDED PI PICO FASTEST SOLDERING COMPETITION ($200 PRIZE POOL) Rules: - MUST PLAY SKRILLEX SCARY MONSTERS AND NICE SPRITES WHILE SOLDERING (loud) - MUST USE A REGULATION PI PICO (Pi Pico 2 and W variants allowed) /cont
English
@mudler_it @ggml @Xiaomi Does it identify birds, insects, mammals, and other living creatures accurately?
English
We built a C++ engine (ced.cpp) that tells you what a sound is, rooster, thunder, a guitar, in ~55 ms on a CPU. No Python in the loop.
It's a from-scratch @ggml port of @Xiaomi's CED audio tagger. Same tags as the PyTorch original, numerically equal , but faster and 2x lighter.
Now you can make your LLM Listen around you!
English
🚨 BREAKING: U.S. government will decide who gets access to GPT-5.6
OpenAI will release GPT-5.6 only in a limited preview to a small group of partners.
Sam Altman told staff the government would be "approving access customer by customer."
Commerce Sec Lutnick personally called Altman warning: don't launch without approvals from other agencies.
A de facto licensing regime.
ITS HAPPENING
English
@blind_via Thank you for the explanation. Looked a bit strange at first, but makes sense now.
English
@MTSlive Dude, straight up plagiarizing my post wtf
NIK@ns123abc
🚨BREAKING: DeepSeek just raised $7.4 billion at $50 billion+ valuation CEO Liang Wenfeng wrote the biggest check himself: $2.8B >Tencent: $1.4B >CATL: $700M >JD, NetEase, IDG Capital: $420M each >China’s National AI Fund: $140M Investors put money into an LP managed by CEO Liang, not DeepSeek >5-year lockup on all shares >external investors get NO voting rights Only exception: the CCP fund invests directly with voting rights and no lockup DeepSeek believes AI should be open-source to actually benefit all humanity.
English
SITUATION EXPLAINED: DeepSeek just raised $7.4 billion at a $50B valuation.
• CEO Liang Wenfeng personally wrote the biggest check of about $3B
• Tencent put in $1.5B, CATL put in $750M, JD, NetEase, and IDG Capital each put in $440M
• All outside investors face a 5-year lockup and get zero voting rights
• The one exception: the Chinese government, which invests directly with voting rights and no lockup, despite contributing only 2% of the round
• Liang's team personally vetted the identity of every limited partner behind the investing funds
@theojaffee: "I think, this is like the first time DeepSeek has ever raised external capital."
MTS@MTSlive
SITUATION DETECTED: DeepSeek has closed its first external funding round, raising more than $7.4B at a $50B+ valuation, per The Information. Investors include Tencent, CATL, JD, NetEase, and IDG Capital. Most investors put money into an LP controlled by CEO Liang Wenfeng rather than DeepSeek directly, keeping control consolidated with its founder. The China National AI Investment Fund is the only investor with direct equity and voting rights.
English
@jholtdigital @grok @milichab It's on the alpha channel, not on stable yet. Run "grok update --alpha" to try it out.
English
English
@prz_chojecki @Kimi_Moonshot @GeminiApp @MiniMax_AI @deepseek_ai @NVIDIAAI Would be interesting to see how Composer 2.5 performs on this.
English
8) If you think it would be useful to cover other models in details, let me know! I'll be rerunning it on more problems.
Also if you need a private benchmarking for your own reasoning model, DM me! We have benchmarks and can help with crafting custom datasets to boost your model's reasoning.
English
While everyone talks about Mythos vs GPT-5.5, we've tested other near SOTA models on our ErdosBench.
Smoke test on 14 problems with 7 models: Kimi K2.6, Gemini 3.1 Pro, GLM 5.1, MiniMax M3, DeepSeek V4 Pro, Nemotron 3 Ultra and Gemma 3 27b.
The winner overall is... Kimi K2.6
English
Recruits, your first prize is here...
A custom GeForce RTX 5080 Founders Edition + PC copy of the game.
Comment #007FirstLightRTX to win 👇
English
