Michael Waldman (@Mwaldman) profile picture

Working on bringing local AI to every phone and laptop. Previously @Meta

Manhattan
Joined August 2026
Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)Photo by Michael Waldman (@Mwaldman)
Michael Waldman retweeted
Another benefit of the shift to open/local that doesn’t get touted enough. Instead of jobs / capital concentrating to a couple hundred employees at closed labs, it will actually distribute to thousands of IT departments.
Recruiters take note: open then local AI will mean the return of the IT department. That will be millions of jobs created 👇🏽
1
4
107
Michael Waldman retweeted
✅ BENCHMARKED: Qwen3.8-27B-pi-adapted 🤗 weights on my 64GB M5 Pro Mac mini. After merging the adapter there is no speed penalty! I also benchmarked Qwen3.8-27B Q8_0 (llama.cpp, thinking mode disabled) with and without the LoRA (median of 3): - Prompt processing speed (4K context): 380 (base) vs. 217 (LoRA) vs. 388 tok/s (merged) - Generation speed (4K context): 9.82 (base) vs. 9.42 (LoRA) vs. 9.97 tok/s (merged) - Time to first token (32K context): 105 (base) vs. 170 (LoRA) vs. 106 s (merged) The merged model (28.6 GB; 37.3 GB peak memory usage) fits entirely in the memory of my Mac mini and is on par with the base model in terms of inference speed. Without merging, applying the LoRA seems to come at a cost of ~40% in prompt processing speed. I also had the base and pi-adapted model solve 12 small Python-related tasks in the pi coding agent (2 times each): Both models solved 22/24 tasks, with the pi-adapted model having a slight edge in terms of solving time (313 s vs. 334 s avg. time per task). ~5.5 hours of results from overnight benchmarks, all on a single Mac mini! Fine-tune open models for your use-case, merge the adapters and enjoy full-speed inference on your desktop.
Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…
15
15
3
134
9,144
Michael Waldman retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
58,688
119,513
24,168
1,142,394
54,084,548
Michael Waldman retweeted
As I’ve said for years, concentration of power, capabilities and wealth is the biggest risk in AI. The natural tendencies for it are super strong and without much much stronger open-source, we might end up in a world with more asymmetry and inequalities than ever!
Anthropic’s IPO could make about 4,250 of its 5,300 employees millionaires, per an analysis by financial platform Hill, including roughly 50 worth more than $100 million each. The estimate assumes a $1.5 trillion valuation, above the roughly $1 trillion Anthropic is now targeting.
193
360
44
3,319
215,499
Michael Waldman retweeted
Why do I never read about Kayleigh from ASU deepseeking state secrets out of the Alibaba CEO? Where is the patriotism? Do we even want to win?
34
91
13
4,164
444,803
Michael Waldman retweeted
You ever met someone who was IRL benchmaxxed? Great college, great career, successful in life—but when you talk with them, you just can’t detect anything behind the eyes?
45
44
25
1,954
160,148
Michael Waldman retweeted
whats being posed as a groundbreaking forensic investigation is literally public information available on GitHub.
37
81
9
1,580
129,832
Michael Waldman retweeted
this seems like a lot of effort to forensically prove something that nobody is concealing you can just look on github and see that DeepSeek regularly publishes Blackwell training/inference kernels
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.
8
22
1
895
41,673
Michael Waldman retweeted
Don't make the mistake of thinking this is about his infidelity. These articles will not stop until he agrees to follow Anthropic's proposal to block inference for unsigned and registered open-models in the firmware of modern GPUs. This is a pretty transparent concerted effort at centralizing control over compute ressources and the ability to execute ever more powerful local models. Jensen is the tip of the spear pushing back at this (in part for his own economic reasons). I'm team Jensen, his affairs are none of my business, not even the alleged support of circumvention of export controls are my business, his position on open-weight models is.
Jensen Huang must go We argue in @TheEconomist that his position is untenable "Like many powerful leaders, he came to believe that he was too important for the normal rules to apply to him." But he is not the only one. Howard Lutnick has failed to stem the flow of chips to China. More heads should roll
404
1,607
147
9,642
294,428
Michael Waldman retweeted
Wow. Fable 5.2 is incredible. It built this random bullshit, check it out! This post is not paid for in any way
75
82
6
2,531
143,461
Michael Waldman retweeted
why are all the smartest 18-23yo stanford/harvard/mit people such losers mindset about everything discuss any idea and they'll by default talk in great detail what things might not work, instead figuring out how to make it work
303
130
86
4,993
493,197
Michael Waldman retweeted
GPT-4 level intelligence has gotten roughly 1800x smaller
50
129
18
2,711
125,179
Michael Waldman retweeted
friends wear friends’ merch
5
1
1
21
3,923
Michael Waldman retweeted
“Getting it manufactured still required several rounds of review and iteration.” Design costs going to zero will make demand for profitable high mix, high vol production go to the moon Valuations for manufacturers will follow. Simple supply and demand.
Replying to @OpenAI
Two OpenAI team members worked with GPT-6 Astra to design a prototype device for talking to Codex in under a week, as a side project. Astra designed the enclosure, circuit board, interface, and software. Getting it manufactured still required several rounds of review and iteration.
10
18
5
380
54,699
Michael Waldman retweeted
Zuck this morning.
262
2,936
198
46,632
700,678
Michael Waldman retweeted
Anyone else just get this Claude update?
75
455
40
11,853
307,938
Michael Waldman retweeted
Claude / Chat will be $2,000+ per month to get the full capabilities. In college Ubering home from the bar literally cost $2-3 and now that ride is $50-60. The VC money ultimately demands a return.
JUST IN: Replit CEO Amjad Masad warns users of subsidized AI subscriptions to "get ready to get rug pulled" as companies face unsustainable operating costs.
179
477
63
10,327
476,024
Michael Waldman retweeted
Own your intelligence. Anyone that opposes this has something to sell you (perhaps your own intelligence back to you.) Local models need to continue proliferating. The “absolute unit” massive models may be out of reach for now locally, so your best option is to, indeed, keep your options open. Own your harness, data, connections, runtime: your agents. Embrace open source.
397
392
182
3,890
283,622
Michael Waldman retweeted
Yann LeCun was right & nobody in this industry wants to say it out loud. 109 million parameters beat 72 billion. He has spent years saying scale is not the thing that makes a model understand the world. Meta bet the other way. He left in November and started his own lab in Paris. Last Thursday he published the proof. The theory everyone has been running on since 2024 is that if you train models big enough, on text, on images, on sound, they all end up with the same picture of reality inside them. Size does it. Nobody ever proved that part. Randall Balestriero and LeCun proved it. And the proof says size was never doing the work. What does the work is which two things you show a model together. A photo and the caption of that photo are two views of one thing. Models trained on pairs like that get there fast. Two paragraphs pulled from the same document are not two views of one thing. Models trained on those get there slowly and expensively. Next word prediction is the second kind. That is what every chatbot in the world runs on. In their tests the first method reached the same understanding after 21.6 million examples. Next word prediction needed 900 million. And a 109 million parameter text model ended up closer to an image model's view of the world than a 72 billion parameter language model was. 660 times smaller.
268
919
161
6,875
810,945
Michael Waldman retweeted
you’re hiding RAM under the floor board aren’t you?
OpenAI CEO Sam Altman believes people owning RAM is wasteful, and that it belongs in data centers instead. "From a global allocation perspective, actually, it’s quite terrible to have everyone have a bunch of the world’s limited supply of DRAM sitting in their own computer."
90
2,737
56
44,110
572,219