Will

@realSharonZhou @AnushElangovan Thanks!

codegen is cheap now. performance isn’t: most generated kernels are kinda mid. iteration and feedback are missing for both the agent and RL env layers. so we're open-sourcing Apex: an end-to-end agent using Claude Code + Codex to effectively optimize AMD kernels, instead of one-shotting them github.com/AMD-AGI/Apex

@0xSero love to help!

In 72 hours I got over 100k of value 1. Lambda gave me 5000$ credits in compute 2. Nvidia offered me 8x H100s on the cloud (20$/h) idk for how long but assuming 2 weeks that'd be 5000$~ 3. TNG technology offered me 2 weeks of B200s which is something like 12000$ in compute 4. A kind person offered me 100k in GCP credits (enough to train a 27B if you do it right) 5. Framework offered to mail me a desktop computer 6. We got 14,000$ in donations which will go to buying 2x RTX Pro 6000s (bringing me up to 384GB VRAM) 7. I got over 6M impressions which based on my RPM would be 1500$ over my 500$~ usual per pay period 8. I have gained 17,000~ followers, over doubling my follower count 9. 17 subscribers on X + 700 on youtube. The total value of all this approaches at minimum 50,000$~ and closer to 150,000$ if I leverage it all. --------------------- What I'll be doing with all this: Eric is an incredibly driven researcher I have been bouncing ideas off of over the last month. Him and I have been tackling the idea of getting massive models to fit on relatively cheap memory. The idea is taking advantage of different forms of memory, in combination with expert saliency scoring, to offload specific expert groupings to different memory tiers. For the MoEs I've tested over my entire AI session history about 37.5% of the model is responsible for 95% of token routing. So we can offload 62.5% of an LLM onto SSD/NVMe/CPU/Cheap VRAM this should theoretically result in minimal latency added if we can select the right experts. We can combine this with paged swapping to further accelerate the prompt processing, if done right we are looking at very very decent performance for massive unquantisation & unpruned LLMs. You can get DeepSeek-v3.2-speciale at full intelligence with decent tokens/s as long as you have enough vram to host the core 20-40% of the model and enough ram or SSD to host the rest. Add quantisation to the mix and you can basically have decent speeds and intelligence with just 5-10% of the model's size in vram (+ you need some for context) The funds will be used to push this to it's limits. ----------------- There's also tons of research that you can quantise a model drastically, then distill from the original BF16 or make a LoRA to align it back to the original mostly. This will be added to the pipeline too. ------------------ All this will be built out here: github.com/0xSero/moe-com… you will be able to take any MoE and shove it in here, and with only 24GB and enough RAM/NVMe to compress it down. it'll be slow as hell but it will work with little tinkering. ------------------ Lastly I will be looking into either a full training run from scratch -> or just post-training on an open AMERICAN base model - a research model - an openclaw/nanoclaw/hermes model - a browser-use model To prove that this can be done. -------------------- I will be bad at all of it, and doubt I will get beyond the best small models from 6 months ago, but I want to prove it's no boogeyman impossible task to everyone who says otherwise. -------------------- By the end of the year: 1. I will have 1 model I trained in some capacity be on the top 5 at either pinchbench, browseruse, or research. 2. My github will have a master repo which combines all my work into reusable generalised scripts to help you do that same. 3. The largest public comparative dataset for all MoE quantisations, prunes, benchmarks, costs, hardware requirements. -------------------------- A lot of this will be lead by Eric, who I will tag in the next post. I want to say thank you to everyone who has supported me, I have gotten a lot of comments stating: 1. I'm crazy, stupid, or both 2. I'm wasting my time, no one cares about this 3. This is not a real issue I believe the amount of interest and support I've received says it all. donate.sybilsolutions.ai

@a1zhang @turbo_xo_ interesting

@turbo_xo_ post training is sufficient. pre-training / mid-training is necessary in the long run for very powerful systems and naturally make better decisions

soon all benchmarks of this form will be obsolete :)

@_LuoFuli Let's goooo

MiMo-V2-Pro & Omni & TTS is out. Our first full-stack model family built truly for the Agent era. I call this a quiet ambush — not because we planned it, but because the shift from Chat to Agent paradigm happened so fast, even we barely believed it. Somewhere in between was a process that was thrilling, painful, and fascinating all at once. The 1T base model started training months ago. The original goal was long-context reasoning efficiency. Hybrid Attention carries real innovation, without overreaching — and it turns out to be exactly the right foundation for the Agent era. 1M context window. MTP inference for ultra-low latency and cost. These architectural decisions weren't trendy. They were a structural advantage we built before we needed it. What changed everything was experiencing a complex agentic scaffold — what I'd call orchestrated Context — for the first time. I was shocked on day one. I tried to convince the team to use it. That didn't work. So I gave a hard mandate: anyone on MiMo Team with fewer than 100 conversations tomorrow can quit. It worked. Once the team's imagination was ignited by what agentic systems could do, that imagination converted directly into research velocity. People ask why we move so fast. I saw it firsthand building DeepSeek R1. My honest summary: — Backbone and Infra research has long cycles. You need strategic conviction a year before it pays off. — Posttrain agility is a different muscle: product intuition driving evaluation, iteration cycles compressed, paradigm shifts caught early. — And the constant: curiosity, sharp technical instinct, decisive execution, full commitment — and something that's easy to underestimate: a genuine love for the world you're building for. We will open-source — when the models are stable enough to deserve it. From Beijing, very late, not quite awake.

@HaochengXiUCB @_akhaliq Interesting

Thanks for sharing our newest work @_akhaliq ! Classic algorithms like K-Means deserve to be revisited in the era of massive datasets and GPUs. Flash-KMeans rethinks the algorithm from a systems perspective to make exact K-Means fast and memory-efficient on modern hardware.

If you're using Claude Code for research: stop making it read directly from PDFs We've introduced a SKILL.md that fetches structured, AI-friendly paper overviews from alphaXiv 👀

CPU CPU CPU CPU

@a1zhang 1.1m writing kernels is insane

Registration link: luma.com/cqq4mojz

$1,100,000 in cash to write GPU kernels for the latest @GPU_MODE competition holy fuck bro

Boomer management doesn't understand that better models lead to more DAU. It's over for Alibaba AI ambitions. They should focus on being like Oracle or Amazon and just sell compute. "Alibaba Cloud has changed the Qwen team's performance evaluation method, using metrics like daily active users (DAU) typical of consumer-grade apps to assess the foundational model team."

@natjin What do you think about prof Jiang?

icymi: this guy has been blowing up on american social media / among the educated elites re: radical theories. he's 'famous' for predicting that trump would win and that US would go to war with iran. here's an unfettered analysis on his claims what's unfolded, and how he aligns with global political agenda... in case you were curious: perplexity.ai/computer/a/dos…

@Linahuaa 😂

Will coding skills still be relevant in 10-20 years? What useful things are you all building with your slave agent codex army? I had a cracked friend from MIT who coded a bot that analyses data and makes value bets on 3rd division college basketball. He lost money. Haven't heard of anything else he built. Btw, he quit our friendship after I showed him my X account, so I haven't shown my X to anyone since then.

WTF you can now train a 5 million context window 8b model on a single node of 8xH100s ???? Most people don't realize that even on long context pretrained frontier models, most RL post-training is only done on a small fraction of that context. Why? Long context RL is notoriously memory hungry, and requires a sharding strategy called Context Parallelism that takes up an inordinate amount of GPUs. This paper for Together flew under the radar, combines the best of Context Parallelism + Sequence Parallel-style head chunking, to get memory efficient long context. The gains are insane, cutting attention memory footprint up to 87% Authors: @m_ryabinin , @sereghik , Maksim Abraham, @ghadiaravi13

@ShrivuShankar Lit

ok decompiling things with opus is actually very addicting here's apparently how netflix and slack work - gist.github.com/sshh12/dda3a89… - gist.github.com/sshh12/4cca8d6… idk maybe wont be that hard to vibe code these after all > use chrome dev tools to explore and provided a grounded tear down analysis of how the site works

@awnihannun what are you building next?

Today is my last day at Apple. Building MLX with our amazing team and community has been an absolute pleasure. It's still early days for AI on Apple silicon. Apple makes the best consumer hardware on the planet. There's so much potential for it to be the leading platform for AI. And I'm confident MLX will continue to have a big role in that. To the future: MLX remains in the exceptionally capable hands of our team including @angeloskath, @zcbenz, @DiganiJagrit, @NasFilippova, @trebolloc (and others not on X). Follow them or @shshnkp for future updates.

literally the most useful update

@BuildWithRakesh @garrytan more like acceptance and no judgement

@garrytan We crave AI companions not because we want artificial relationships. Because reliable human attention became scarce.

This is not the happy demo path, Apple or Google would never make this one of their launch videos... it's not what you will hear about in a TED talk, but it's real AI: Doesn't get tired, doesn't ghost A lot to think about in this one

@TeksEdge what device is that

Not only does Qwen3.5-35B-A3B decode @ 30tps on DGX-Spark, but the smaller flavor 27B scores near SOTA level on Humanity's Last Exam! Imagine running on a small portable PC!

@LandseerEnga need this!

Agents can one shot mobile apps, but testing is still the bottleneck. So we built a CLI that gives Claude Code the one thing it was missing - eyes and hands The best part? t's fully vision-based: - No scripts, no selectors, no element IDs - It interacts with the app exactly like a human would Claude now writes code → tests it on the app → sees what broke → fixes it Spawn a cloud agent, go to sleep, and wake up knowing it actually worked.

Been using it non stop since sat Got 10 codex, 10 Claude code, and 5 Browser tabs open rn These guys are cracked!

!!!!!