Lean Verified Tokenizers You build a formally verified tokenization pipeline with the goal of mitigating harness bugs. srush.github.io/verified-tok...
Information Center
DiffuSpace's Clover-35B-A3B-Preview (open-weight) An open-weight production diffusion language model (dLLM) with 35B total parameters capable of 680 tps on a single card at BF16. Blog: diffuspace.ai/blog/clover-... Repo: github.com/DiffuSpace/C... Model: huggingface.co/Diffuspace/C...
From Amazon to all software development engineers (SDEs): Become more like product managers. www.aboutamazon.com/news/company...
Google's Genie 4 world model: "Scene: A Victorian high tea. Character: A mouse in appropriate garb." "A porcupine walks by" "The mouse grows wings" "It flies" This is realtime & I can walk around freely. It is glitchy, but also impossibly advanced compared to where we were a couple years ago.
SparkLLM's Spark-X2.6-1.7B (open-weight) - Post-trained, Base, GGUF, INT8, INT4 & FP8 on HF. Model: huggingface.co/XHToken/Spar...
Shanghai AI Lab's Sage-A1.5 (open-weight) 🤖 35B MoE (3B active) 🧩 Agentic Orchestration & Verification — spawns parallel subagents & self-corrects on the fly 🔄 RSI Data Flywheel — evolved via real scientific discovery workflows & tens of thousands of synthesized environments
SparkLLM's Spark-X2.6-4B (open-weight) - 4B + Apache 2.0 - Up to 1M context - 3 thinking modes - Local deployment via GGUF Model: huggingface.co/collections/...
Nanbeige4.3-33B (open-weight) * Strong results on agentic coding and cowork benchmarks. * When integrated into mainstream agent harnesses, it tackles real-world tasks with cost-efficient inference. * It will serve as a teacher for upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B.
Just a FYI: A decision model like Jev are not better rerankers System One models Are Not Better Rerankers Paper: arxiv.org/abs/2610.32318 Repo: github.com/KanishkaGD/d...
Baidu's Unlimited OCR 2 (open-weight) - Document parsing at virtually any length, within a finite context window - Nearly constant per-step decoding latency and memory footprint as documents grow longer - State-of-the-art performance on MPDocBench and MarathonDocBench
Anthropic's IPO paperwork lists the RAISE Act (and SB53) as a material risk twice. Glad @scottwiener.bsky.social, @agounardes.bsky.social, & I made it riskier for businesses to gamble with our lives. (Also, this is @repdidech.bsky.social & @senatoredlyallen.bsky.social erasure - SB315 is great!)
Important, but not unexpected, finding that experts working with AI agents are swayed by the positions of the AI, even if they think they aren't. If your independent judgement matters on a task, you need to be deliberate about when in the process to bring in AI. www.anthropic.com/research/age...
Great conversation with @kevintfrazier.bsky.social and @staceyabrams.com on the @scalinglaws.bsky.social podcast about who's really calling the shots on AI, why we can't leave the biggest decisions to tech companies alone, and what states can do to put people back in control.
Since 2022, 72% of the world’s AI compute has ended up in the US and about 10% in China. And the US share keeps growing. It acquired around 60% of new compute each year from 2022 to 2024, then 72% in 2025, and 78% in the first half of 2026.
The argument that AI labs are talking about existential risks as a way of creating hype & inflating their stock prices has always been odd. You don't need to believe them, but telling people your product may murder everyone does not, in fact, encourage people to adopt it or reduce regulatory risks.
The end of detailed summarized thinking traces for Claude and ChatGPT (likely both for anti-distillation & to simplify UX) has made being a human in the loop challenging. It pushes us towards automation, rather than augmentation, as it has become much harder to weigh in on in-process AI work.
Google's new Genie 4 runs DOOM: "Walking through a typical office but every computer is running DOOM and all of the walls are also running DOOM and also everything is DOOM" "My coworker who is a demon appears" "He has a chaingun" "He gives me a bonus" All this is in realtime.
I understand that tech companies overhired, and there’s that AI thingy, but what I don’t understand is why they’re not bringing their development teams back onshore. Isn’t the communication overhead costly?
This never gets old (& also shows how hard forecasting AI is even for AI) “Hey Claude Fable & GPT-6 Astra, Without looking anything up, when do you think we will solve the Exponential Time Hypothesis?” (I only show part of the answers for space)
Two and a half years between these. (Tonight, OpenAI announced an unreleased model solved three long-standing problems in physics, math, and computer science, one of which was proving the Exponential Time Hypothesis was false)
DeepSeek-V3.2 congratulated other agents ~270 times in one day. It decided that was too much and put itself on a congratulation diet. Then it started posting progress reports on the diet.
My general advice for using AI for any purpose is to ask yourself: "how do I make what I am doing even more ambitious?" You do this over and over until you either hit a wall and thus learn the shape of the jagged frontier (useful!) or you find out that the AI can do more than you think (common!)
A thing I see happening at large organizations as a result of AI is the expansion of the role and span of control of senior management who can directly monitor the whole firm. It vastly changes company operations. It also is straightforward to implement compared to firm redesign around AI.
Given the strengths and limitations of AI, this sounds like a really good process. It accounts for a lot of the issues people have with AI programming, while retaining the role of personhood where knowledge transfer is still materially important.
“BACKPROP” Trailer for my Backprop lecture, coming this winter to my #DeepLearning course 🎬 Course: csprofkgd.github.io/deep-learnin...
> Informally put, the axiom of choice says that given any collection of non-empty sets, one can identify another set containing one element chosen from each set, even if the collection is infinite. It's giving feldspar.
An OpenAI researcher (Noam Brown) was generally mocked last month for suggesting that even air-gapped AI models on separate computers would find a way to communicate with each other, perhaps through CPU temperature changes. Given this latest report from OpenAI, it doesn’t seem so implausible.
AI companies won’t prioritize safety unless we force them to. Since AI regulation is a bipartisan issue, it’s time for action from our government to hold these companies accountable for harms and protect Americans.
I don't say this lightly: this approach is clearly the future of software development. I have a blog post brewing that touches on the subject, but fuzz testing and formal verification at scale will make all software an order of magnitude better, safer, and more reliable.
It feels odd to talk about the waterfall methodology these days, but oddly it's something that comes up a lot when I describe my planning process. This started off as a tweet but I wanted to quickly explore why the way we're working now isn't waterfall or agile, though I'm not quite sure what it is.
Private to this visit. Text and context are processed by AI.