
šØāš» Senior AI Research Engineer @NVIDIA. Ex-Groq. Occasional angel investor. Opinions are my own.
Joined July 2009
- Tweets4.5K
- Following423
- Followers6.3K
- Likes10.8K
You canāt prove AI hasnāt been used unless you can prove you did all the work before approx 2024-2025
This is massive. A story I've been following for years. He thought he had more time but didn't. Alessandro Masullo announces a separate, purely human proof of the Riemann hypothesis. math216.wordpress.com/2026/1ā¦
1
258
Rick Lamers retweeted
(1/4) why do image models chop every picture into the same square patches? š¤ (hint: they shouldn't)
following the bitter lesson, we let diffusion models *choose* how to tokenize an image. it spends fewer tokens on the boring background, and more tokens on our faces! š§µ
>You have essentially infinite training data, an un-hackable reward
Swarms banging their heads on compression will continue to pay dividends for quite a while
There has been some rather impressive progress on the Hutter prize for text compression recently (~10%; a huge trend break). Interestingly, it's all based on the idea of training / overfitting a small transformer on the data to be compressed. This model will assign high likelihood to the data, and consequently arithmetic coding (using said model) will assign short code lengths to the tokens. Because the model itself is not known by the receiver, its weights must be compressed and included in the code length as well, and so you train with a loss that includes that cost as well.
All the way back in 2020/2021 when I was working on generative models for compression, we explored this idea for video compression: take a pre-trained video compression model (discrete-latent AE), and overfit it on one video using a loss L = coding cost of weight-delta + coding cost of data given model (i.e. log likelihood). We used a spike-and slab prior on the weight delta so that it is cheap to encode "no change" to the pre-trained weights.
We found it to be very effective, but needing to train a model at encoding time + inferencing it at decoding time is kind of a hassle, so we didn't pursue it further. It still feels like a long shot, but given how competitive such approaches now appear to be, and how robust (quantization-aware) training techniques have gotten, it may be worth another look. Especially for cases where encoding compute is not a major concern but code length is (i.e. content on any platform with lots of users, where you encode once and transmit + decode many times).
Incidentally, compression is also an excellent task for research on coding / AI research agents. You have essentially infinite training data, an un-hackable reward, and according to Kolmogorov / Chaitin it will never saturate.
References
@tivaro @IamHuijben Cohen, Overfitting for Fun and Profit: Instance-Adaptive Data Compression arxiv.org/pdf/2101.08687
@OriYoran The KoLMogorov Test: Compression by Code Generation arxiv.org/abs/2503.13992
Hutter Prize (Marcus Hutter, @mhutter42): prize.hutter1.net/
Matt Mahoney, Large Text Compression Benchmark: mattmahoney.net/dc/text.html
Hutter prize SOTA over time
epoch.ai/data/hutter-prize?vā¦
Recent entries
fx2-cmix-transformer, Vladimir Ivanov (the award winner; offline-trained transformer, 4-bit weights shipped): github.com/astOwOlfo/fx2-cmiā¦
zmix, James Byrne (adds the coded size of the weights to the training loss): zmix.frostbyrne.io/v1.0/writā¦
cmix-neif-beehive, Vladimir Malinovskii (differentiable rate term on the weights, with ablation): github.com/galqiwi/cmix-neifā¦
cmix-neif 2.0, Marcus Wƶlk (best entry so far; entropy-coded weights with bits-back): github.com/marwoelk/cmix-neiā¦
tufazip @tufalabs, Tommy He / Tufa Labs (shipped prior plus online adaptation; analysis of model bits vs. data bits): tufalabs.ai/research/tufazipā¦
DNA-4, Ao, Sun & Sun (RWKV-7; arithmetic-coded 4-bit weights, then online fine-tuning): github.com/ziq-ao/DNA4-RWKV
1
236
We need ARC-AGI-4
GPT-6.2 Astra from @OpenAI on ARC-AGI (Verified):
- ARC-AGI-3: 100.0% (SOTA), $4.1K (provider adapter), 70.3%, $24.2K (standard)
- ARC-AGI-2: 95.8% (SOTA), $0.50/task
- ARC-AGI-1: 98.5%, $0.17/task
GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3.
1
1
338
Rick Lamers retweeted
Which way, Western models?
Replying to @EpochAIResearch
We also evaluated how many serial reasoning steps GPT-6.2 Astra can do in a single output token (without chain of thought). It can perform 12.3 steps, versus 10.5 for GPT-6 Astra and 8.6 for GPT-6.1 Sol, the previous highs. We measure the number of steps models can perform at an accuracy of 25% above chance, where each step is normalized for complexity.
Before GPT-6 Astraās release, the most capable model was Claude Fable 5.1, which is able to perform only 4.2 serial steps. The oldest model we assessed is Claude Opus 3, which performed 2.3 steps.
This capability is notable because models that can do complex reasoning without any chain of thought may be harder to monitor.
We hope to finalize this and other measurements of more abstract model capabilities, and release them in the coming weeks.
Also one-shotted a neat Mac utility for this model to use it locally on Apple Metal! MIT licensed ofc
Still on Whisper large/turbo for local English ASR? Rumble-2-Flash (0.5B, Apache 2.0):
vs large-v3: 67% smaller (1.02 vs 3.09 GB), ~30% fewer errors (~5% vs ~7.2% WER)
vs turbo: 37% smaller (1.02 vs 1.62 GB), 31% fewer errors (5.9% vs 8.6%)
AMI meetings: 9% vs 16%, -44%
Thomas/Raoul/team cooked!
Rick Lamers retweeted
Over the past week, Iāve 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 lifeās 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 Boardās independent investigation.
Since my departure, Iāve seen an outpouring of calls from NVIDIAās 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
Rick Lamers retweeted
Wow that escalated quickly. Was gone for a few weeks and now pretty much everyone launched their own decision model copycats:
- Sep 23: Together AI, Tev1-4B
- Sep 24: Alibaba Cloud, decision-model
- Sep 29: Liquid AI, d1
- Sep 30: Databricks, ai_decide
- Oct 1: Amazon AWS, Strands Decider 2B
- Oct 1: Cloudflare, Clef and Clef-flash (Clef-omni on Oct 9)
- Oct 1: Perplexity, Decider (v1.3 on Nov 12)
- Oct 6: OpenAI, Decisions API (GPT-6 Luna)
- Oct 7: H2O.ai, H2O-Lightning-4B
- Oct 7: Snowflake, Snowflake Decision
- Oct 9: Microsoft, Microsoft-Decision-1
- Oct 15: IBM, Granite Switch 4.1 with Decision adapter
- Oct 21: Baseten, Mammoth-1 and Elephant-1
- Nov 6: Salesforce, Koa Decisions
- Nov 11: Google DeepMind, Gemini Decision 1
According to a Vercel report, 26% of requests are now routed through decision models š
From the reply:
> In case that wasnāt clear, the process is:
> - Generate an AI image of a living room
> - Add a real photo of a car in Photoshop
> - Display the altered image on a laptop
> - Take a real photo of the laptop
> - Upload that photo to Pangram
That's wild š«” @pangram
Rick Lamers retweeted
Each agent gets a skill library, motion planner, simulator & 10 demos. It writes skills, simulates ~30x, then asks for ~10 real trials
6 frontier models
5,400 real-world rollouts
Spatial generalization measured
Results: never-before-seen success rates and new behaviors! 2/12
Rick Lamers retweeted
ā¤ļø Nvidia Colleagues
ā¤ļø Jensen
ā¤ļø AI advancements
Mission is the boss
Rick Lamers retweeted
Inkling 2 Smallās weights are now available for download on HuggingFace under an Apache 2.0 license.
Make Inkling your own!
huggingface.co/thinkingmachiā¦
a definition of AGI I hadn't seen yet (maybe a me issue) but enjoyed more than expected
> I believe weāre in the early stages of the singularity, defined as the period when the rate of technological and scientific progress exceeds the rate at which we humans can parse out and make sense of it
src: @jackclarkSF
1
278
JIT your Rust folks (esp during agentic dev!)
Rust has compile time issues. I have a fully optimized LLM accessible flow with Relax, a Rust JIT. Now with fable 5.2 I can finally have a worthy website for this absolutely jawdropping piece of technology made by machines, for machines. makepad.github.io/relax/
275





























