Mike B (@maik18) header
Mike B (@maik18) profile picture

Co-founder/CTO @turintechai. Building Artemis: evolutionary search for faster AI and code.

London, England
Joined April 2009
Photo by Mike B (@maik18)Photo by Mike B (@maik18)Photo by Mike B (@maik18)
Pinned Tweet
Do you really need more RAM to run AI locally? NVIDIA's Nemotron 3.5 Bolt needs ~19 GB. Our 3 GB version runs in 5.7 GB, ~3x faster, and still gets tool calls right. How much RAM does your local setup have?
3
1
2
67
We think these are the closest to the original you'll find at 3–4 GB. There's a lot of great work out there, though, so if you know a better file at any of these sizes, send it and we'll run the same test. Thanks to Unsloth, bartowski and the llama.cpp community for the work this builds on.
1
21
Serving on GPUs instead? FP8, W8A8, W4A16 and NVFP4 checkpoints for vLLM are up too, 23–47% faster than the original on one RTX 5090.
13
Mike B retweeted
this guy figured out how to leak your exact location from any video with sound using the Doppler curves of nearby airplanes
588
1,895
648
34,983
3,724,308
Artemis, our optimisation platform, discovered the compression method. Thanks to @Cloudflare, @perplexity_ai and @jaredpalmer for the original models. Independent conversions, not affiliated or endorsed. Which decision model should we shrink next? huggingface.co/collections/t…
29
vs the usual 4-bit (Q4_K_M) downloads: 16-25% smaller, within about 2 points of agreement. At about the same size as the public IQ3_M files, ours keep ~95% of Clef and Clef-Flash's decisions vs 88-90%. For the Decider, the only public files that run on stock llama.cpp today.
8
Five decision models, file size and how often each gives the original's answer (2,000 test decisions): Decider 27B: 53.8 → 15.3 GB, 96.8% Clef 27B: 54.1 → 14.8 GB, 95.6% Clef-Flash 9B: 18.2 → 4.9 GB, 94.1% Kev-9B: 17.9 → 4.8 GB, 92.2% Kev-4B: 8.4 → 2.5 GB, 93.0%
12
We compressed Cloudflare's Clef and Perplexity's Decider to run in about a third of the GPU memory, and they still give the original's answer ~96% of the time. Decision models, now on your own machine. Free on @huggingface, stock llama.cpp. 🧵
2
45
Once you get Astra access run this prompt against your codebase (or just a part of it). It's magic. "Think from first principles about what we're trying to achieve here. Interrogate what you built before calling it done: 1. is anything here unnecessary, overly complicated, or based on weak assumptions? Challenge them. 2. What can be deleted entirely? 3. What can be simplified now that unnecessary pieces are gone? Then make the changes. Prefer deleting over simplifying, simplifying over optimizing, and optimizing over automating. It might be done too - you dont HAVE to go and make changes. if its good, leave it alone"
78
214
22
5,212
403,417
Mike B retweeted
🚨The mysterious model on OpenRouter is insanely good. Ox Alpha is reportedly destroying people’s internal benchmarks and one-shotting coding tasks normally done by Fable 5. It is currently free, multimodal, has a 1M-token context window, and OpenCode says it has capacity to serve 100T free tokens a day in total. Nobody knows who is behind it. There is clearly serious compute, and probably a major AI lab behind it. What model do you think is behind it?
132
63
28
1,981
226,058
Mike B retweeted
Met a kid y'day who has 50+ agents running with a harness tuned for cyber security and a search index of CVEs with a QLoRA'd GLM 5.2 and tries to break production websites 24/7. Didn't go to college. Makes money in bug bounties. The kids are definitely alright.
159
228
26
6,305
400,391
OpenAI merged our fix into Whisper: one file, seven lines, 4.9× on the kernel it repairs, all 200 test transcripts byte-identical. This weekend, one layer up, WhisperLiveKit merged another: ~⅔ of ASR compute gone, accuracy unchanged. Same stack, two layers. Found by Artemis
4
2
254
Why care: wasted compute is wasted energy and hardware, multiplied across a codebase downloaded millions of times a month. Seven lines, and the savings ship everywhere. Software optimised by AI for its hardware; hardware, increasingly, designed by AI for its software. Optimisation becomes a property the stack has. How we found it: turintech.ai/blog/fixing-a-s…
11
The WhisperLiveKit finding: the live server re-encoded audio it had already seen — 40 encoder passes where 10 suffice. Coalescing them cuts ASR compute ~65% on CPU. The honest trade: first-word latency +1.4s. So it ships opt-in, off by default — a dial instead of an accident. github.com/QuentinFuxa/Whisp…
11
The Whisper bug: under beam search — the CLI default — the flash-attention kernel silently declines the cross-attention work and falls back to the slow path. No error, no warning, on every run since Whisper adopted SDPA. Correct answers hide slow paths. Your logs won't tell you. github.com/openai/whisper/pu…
32
Mike B retweeted
Highly-recommended overview of self-improving agentic systems. (bookmark it) Self-improving agents are moving from research demos into deployed systems. This survey frames a modern agent as a foundation model coupled with an operational scaffold, then formalizes self-improvement as a self-induced update operator that commits changes to either. Paper: arxiv.org/abs/2607.13104 Learn to build effective AI agents in our academy: academy.dair.ai/
41
149
9
798
47,288