1.
pg_gpu: A GPU-first population genetics library (t.co)
Population genetics gets a full-stack GPU library with a bioRxiv paper and an implementation aimed at making large-scale analyses faster
Backlist — 31 May 2026 UTC
Capped agent/coding items in favor of science, security, hardware, markets, robotics, and design despite many strong AI-coding candidates
2.
BIS says AI chip exports to China-headquartered firms abroad need licenses
A licensing loophole that let cutting-edge U.S. AI chips flow to overseas subsidiaries of Chinese companies is being explicitly closed
3.
UniLab speeds up robot reinforcement learning with a CPU+GPU pipeline
Tsinghua AIR reports a 3–10x end-to-end wall-clock improvement for robotic RL by splitting simulation and learning across heterogeneous hardware
4.
CoreWeave and Dell report Rubin VR200 NVL72 racks passing L11 diagnostics
The first public Rubin rack bring-up signal marks a transition from NVIDIA roadmap slides to cloud-level software integration work
5.
A new paper attacks the stability of Walrasian market equilibrium
A 150-year-old question about whether market equilibria converge or blow up is getting a modern formal treatment
6.
WP Maps Pro bug exploited to create WordPress admin accounts (t.co)
An actively exploited plugin vulnerability is granting attackers administrator access on WordPress sites
7.
GPT-5.5 vs Opus 4.8 on DeepSWE: score, latency, and cost
DeepSWE results put model quality, speed, and price in the same frame instead of treating coding benchmarks as a single leaderboard
8.
NanoFold: Protein-structure training data for independent researchers
A Hugging Face dataset of 10k sampled chains lowers the entry barrier for training protein structure prediction models outside large labs
9.
Reaction Engines' SABRE precooler was a real propulsion breakthrough
Cooling Mach-5 intake air fast enough to protect a lightweight engine remains one of the hardest heat-transfer problems in aerospace
10.
ReActor retargets human motion capture to robot bodies
Disney Research tackles how to transfer motions across robots with different joints, shapes, and mass distributions
11.
The biggest VC rounds have underperformed the S&P 500
A portfolio of the ten most heavily funded venture-backed companies would be down 120.5% relative to the index, challenging the idea that capital raised tracks importance
12.
parakeet.cpp: GGML-based Parakeet inference (t.co)
A local GGML pipeline reportedly runs Parakeet speech inference twice as fast as an ONNX setup on Apple Silicon
13.
A gentle introduction to lattice-based cryptography (t.co)
A new IACR preprint offers an accessible route into one of the main families behind post-quantum cryptography
14.
OpenLogi: A Rust, local-first replacement for Logitech Options+
HID++ mouse configuration gets an open native app with no account requirement or telemetry
15.
TCOD: Temporal Curriculum On-Policy Distillation (t.co)
Tongyi Lab’s method adds a temporal curriculum to on-policy distillation, a post-training direction that has been gaining traction
16.
RoboOS: A shared-memory operating layer for multi-robot systems (t.co)
A hierarchical embodied OS points toward coordinating robot fleets through shared memory and composable system layers
17.
Retatrutide's liver-fat data
The cited retatrutide data shows an 86% liver-fat reduction at 48 weeks, far beyond the reductions cited for tirzepatide
18.
Neobanks are being squeezed from bank charters and stablecoins
Fintech margins are being compressed from above by charter-seeking platforms and from below by public rails and stablecoins
19.
Why the ATM didn't kill bank tellers (x.com)
Automation often changes the economics and scale of a job before it eliminates the job, making the ATM example newly relevant for AI labor debates
20.
WindBorne's planetary balloon constellation
A team under 100 people is operating a global network of weather balloons as live infrastructure, with each dot on the map representing hardware in the sky
21.
A one-person watch brand builds a titanium piece unique
The process from sketch to finished watch is a durable artifact of small-scale precision manufacturing and industrial design
22.
halftone-dots: Open-source image morphing effect (t.co)
The repo turns a popular visual effect into a reusable tool for generating and morphing halftone dot images
23.
mrdoob's WebGPU, SSGI, and MediaPipe experiment (t.co)
Browser graphics, screen-space global illumination, and body tracking combine into a compact interactive WebGPU demo
24.
FlashAR claims 22x faster autoregressive image generation
A lightweight vertical prediction head enables parallel decoding for existing autoregressive image models without retraining from scratch
25.
Paint.NET finally acquired paint.net
After 22 years of using getpaint.net, the creator recovered the obvious domain and closed a long-running piece of software history
26.
Sine-wave inputs improve periodic gait generation in MLPs (t.co)
Injecting sinusoidal signals into an MLP improves convergence on periodic robot gait generation, a small architectural trick with practical control implications
27.
NVIDIA SkillSpector scans agent skills before installation (t.co)
AI agent skills are becoming a supply-chain surface, and static plus semantic checks target prompt injection, credential theft, and data exfiltration before deployment
28.
San Francisco's housing shortage exports service workers
Late-night service work increasingly depends on people commuting from Solano and Contra Costa because the city refuses to build enough housing
29.
A banana-controlled projector physics game
A Three.js scene, RF-DETR vision model, and real bananas become physical launchpads in a live computer-vision game
30.
Recreating Apple's breathing status LED
A 2002 Apple patent for a pulsing sleep indicator becomes a hands-on exercise in hardware nostalgia and interface craft
31.
Layout of the latest video I just uploaded.
Layout of the latest video I just uploaded. 01:43 - Dual-sortation (Unit sortation) 02:42 - Custom cardboard box machine 04:05 - The "Air-Only" Vacuum Lifter 05:52 - Quick-repair rack kits 08:06 - The "Taco Turn" Genius 09:56 - Co
32.
Open Code Review is a CLI tool that uses an LLM agent to read Git diffs and produce structured, line-level code r…
Open Code Review is a CLI tool that uses an LLM agent to read Git diffs and produce structured, line-level code review comments. - Combines deterministic engineering with an agent for reliable review quality - Uses smart file bundling and
33.
wanted subagents so we're gonna figure out how to make it work with dotagents (t.co)
wanted subagents so we're gonna figure out how to make it work with dotagents https:// github.com/getsentry/dota gents/pull/106 …
34.
I made a git repo for all my agent settings (t.co)
I made a git repo for all my agent settings Includes: - skills - pi extensions - symlinks to share across all agents https:// github.com/anishthite/age nt-dotfiles/tree/main …
35.
Bug fixes shipping to Grok Build 0.2.13 (release notes will be available in the TUI and on change-log website)
Bug fixes shipping to Grok Build 0.2.13 (release notes will be available in the TUI and on change-log website) We are leveraging the alt-screen to better handle your background tasks, subagents, monitors with smart grouping allowing you to
36.
I cloned the SpaceX site with Grok Build and Firecrawl. (x.com)
I cloned the SpaceX site with Grok Build and Firecrawl. The new design cloner workflow in the @firecrawl CLI packaged the full page and 250+ artifacts into a design,md for the agent to build from. Great starting point for designing off
37.
In the profiling blog post we talk about the CPU chain of dispatch
In the profiling blog post we talk about the CPU chain of dispatch > how operations are wrapped > how to annotate operations > what are the cuda launches > why is there a gap between the dispact and the kernl launch
38.
Btw, I've been doing this with a vllm fork that adds a steering runtime that only shows about 2.7% throughput los…
Btw, I've been doing this with a vllm fork that adds a steering runtime that only shows about 2.7% throughput loss at 32 batch size on Gemma 3 27B. It also adds an activation capture consumer plugin system with the example being a file stor
39.
doesn’t invalidate mythos lower-bound but prop 7.1 seems wrong as written.
doesn’t invalidate mythos lower-bound but prop 7.1 seems wrong as written. it tries to prove u(P)=n^(1+o(1)) needing an upper bound on u(P). but prop. 4.2 gives u(P)/n ≥ ½D(T)e^(−C₂d), a lower bound. so log(u(P)/n)≤log D(T) doesn’t follo
40.
been very interested in pretraining research recently so to get started I reproduced the baseline modded nanogpt …
been very interested in pretraining research recently so to get started I reproduced the baseline modded nanogpt setup and tweaked it to train on a single h100 and reached 3.278 fineweb val loss in 9.37B tokens (~5hrs)
41.
I'm exploring soft muon for RL. It may encourage diversity and exploration while being more robust to noisy small… (t.co)
I'm exploring soft muon for RL. It may encourage diversity and exploration while being more robust to noisy small singular modes of the policy gradient compared to full Muon. https:// nilin.github.io/rl-diversity-s oft-muon/ …
42.
OpenAI's Responses API strikes again
OpenAI's Responses API strikes again We run OpenAI models on Azure and fail over to another region when one gets flaky But Responses item IDs are tied to the region that created them, so the new region can't read messages from the old one
43.
in 2026, no need for an infra hire to give your agents the best env
in 2026, no need for an infra hire to give your agents the best env just run: > box new a few sec later you get a full linux VM in the cloud for agents, with sudo, desktop at 60fps, chrome, ssh, most prog. languages, docker, android emul
44.
post a screenshot of some ai generated css slop: 1.5M views, a gazillion comments
post a screenshot of some ai generated css slop: 1.5M views, a gazillion comments post a long blog post on a shitty robot with a pretty neat fully local speech-to-speech pipeline that becomes a STEM project for a bunch of kids: 15k views
45.
This is the actual bottleneck. The models are smart enough already. What is missing is the company-specific conte…
This is the actual bottleneck. The models are smart enough already. What is missing is the company-specific context locked in senior people heads. Whoever cracks knowledge extraction at the company level unlocks the rest. As you work on t
46.
Finally got SMP working on SMPL.
Finally got SMP working on SMPL. Early 900 epoch clip, tons to improve, but it's solid. Had to divert a bit from the paper so i'll make a longer post with all that when it's had more time to bake. Absolutely stoked.
47.
I wanted to know exactly what the primary sources were saying about the big boom in the Northeast today (which wa…
I wanted to know exactly what the primary sources were saying about the big boom in the Northeast today (which was very loud and startled my cats to knock over all my plants). Claude compiled this little report. Fact checked this a few ti
48.
corollary:
corollary: generative language modeling vs classification of (arbitrarily long bodies of) text as being synthetically generated have the same complexity
49.
Hello,
Hello, you don't need a better RL algorithm. Just cook your sim-learning pipeline.
50.
Humanoid logistics sorters are starting real shifts in both the U.S. and China
Humanoid logistics sorters are starting real shifts in both the U.S. and China Figure is entering Catalyst Brands’ Reno logistics center, inside the retail network behind JCPenney, Aéropostale, and Brooks Brothers. RobotEra’s M7 is worki
51.
We have released Triton puzzles on (x.com)
We have released Triton puzzles on @TensorTonic . We got an insane reaction on the CUDA sheet with more than 1000 submissions. GPUs are completely free, no local setup needed. go run your code and burn those credits
52.
Geometric foundation models, like #VGGT, have the potential to enhance Vision-Language-Action (#VLAs) models. but…
Geometric foundation models, like #VGGT, have the potential to enhance Vision-Language-Action (#VLAs) models. but do they actually help? Intuitively, VGGT-like methods can inject geometric understanding about distances, contacts, etc. that
53.
Met with a founder last week
Met with a founder last week I invested over 4,000 Euros into his Series A last decade "We're finally profitable!" He proudly told me I was skeptical We pulled out his books He was missing a huge expense "Where are the carbon credits
54.
Law of Neural Interaction: Depth-Width Shape, Interaction Efficiency, and Generalization (t.co)
Law of Neural Interaction: Depth-Width Shape, Interaction Efficiency, and Generalization Wenjie Sun, Jinning Yang, Shuai Zhang, Mengnan Du https:// arxiv.org/abs/2605.27989 [𝚌𝚜.𝙻𝙶]
55.
new hobby: give gpt-5.5-low impossible task to grind on all night
new hobby: give gpt-5.5-low impossible task to grind on all night /goal improve performance of xyz benchmark by 100x. each change you land must improve it by at least a 10% with no more than +10 net lines of code doesn't finish but finds
56.
I see now.
I see now. My opinions on Opus 4.8 are NOT valid. Unlike GPT 5.5, is heavily trained on my work. I just asked both models: "without internet, describe the Interaction Calculus" Opus nailed it, character perfect, using my own notation(!!!)
57.
Generative supervision unlocks embodied intelligence
Generative supervision unlocks embodied intelligence Tencent Hunyuan and Tsinghua University release GEM, a VLM that learns physical grounding by predicting depth maps during pre-training, achieving state-of-the-art results on embodied ben
58.
Cool work!
Cool work! A nice reminder that for problems with classical optimization-based solvers, iterative refinement can be very strong (not only inductive bias). Have you guys tried Muon btw?
59.
It's a gorgeous and funny bug.
It's a gorgeous and funny bug. Fwiw, I'm the biggest eBPF fanatic, but I don't think unprivileged users should be able to load arbitrary eBPF programs.
60.
We are arriving in ICRA, Vienna for presenting our Behavior Foundation Model for Humanoid Robots at Interactive S…
We are arriving in ICRA, Vienna for presenting our Behavior Foundation Model for Humanoid Robots at Interactive Session 1 (Hall C), TuI1I.120, June 2nd. Come and strike up a conversation on anything about humanoid robots and beyond~
61.
reading through the exe git log this afternoon and learned that the web terminal is now powered by libghostty.
reading through the exe git log this afternoon and learned that the web terminal is now powered by libghostty.
62.
do NOT run /goal on new benches over the night!!
do NOT run /goal on new benches over the night!!
63.
this morning, i continued a task on codex from kittylitter while taking a piss, then checked in on the training j…
this morning, i continued a task on codex from kittylitter while taking a piss, then checked in on the training jobs i had codex run on modal from a session in pi, then was able to import a session from droid to continue work on a local llm
64.
Gebrauchsgraphik, 04, 1943. Cover design by Leonardo Spreafico (t.co)
Gebrauchsgraphik, 04, 1943. Cover design by Leonardo Spreafico https:// designreviewed.com/artefacts/gebr auchsgraphik-04-1943/ …
65.
Hey (x.com)
Hey @xai @elonmusk — Grok Imagine is fantastic, but the default batch size is burning through SuperGrok quotas much faster than it should. One prompt currently triggers ~12–20 internal renders on the backend. The UI then shows around 1
66.
Shameless plug but this nice work supports our ICLR paper—ICL Activation Alignment—pretty much spot on.
Shameless plug but this nice work supports our ICLR paper—ICL Activation Alignment—pretty much spot on. - Activations (internals) provide a much stronger learning signal than just tokens. - Brings sample efficiency and avoids spurious cor
67.
Witnesses in Key Authorizations on Tempo
Witnesses in Key Authorizations on Tempo This will let us to do 1 Passkey prompt when the user logs in to an application which also doubles down as a auth challenge signature AND a access key authorization Net - you passkey once, get read
68.
Qwen 3.6 has me excited about local LLMs again. Running fully local, slaying bash commands like it's nothing
Qwen 3.6 has me excited about local LLMs again. Running fully local, slaying bash commands like it's nothing
69.
gpt-realtime-2 is genuinely fast, and this demo is great.
gpt-realtime-2 is genuinely fast, and this demo is great. no demo runs long enough to show the ceiling though. we ran it 60 turns. around 5 minutes in it went silent: took our audio, returned zero bytes, no error. then the connection drop
70.
Would be funny if inoculation prompting results in models that are much better at sandbox escapes and other forms…
Would be funny if inoculation prompting results in models that are much better at sandbox escapes and other forms of hacking because they get to spend the whole RL run practicing these things
71.
Can we make a model smarter without post-training? (x.com)
Can we make a model smarter without post-training? Yes, by changing how we sample at inference time ( @aakaran31 @du_yilun ) We make this more practical by sampling more efficiently, in turn, enabling faster reasoning Details by @fel
72.
This week's #PaperILike is "Plan-based Reward Shaping for Reinforcement Learning" (Grzes & Kudenko, 2008). (t.co)
This week's #PaperILike is "Plan-based Reward Shaping for Reinforcement Learning" (Grzes & Kudenko, 2008). A nice combo of planning and RL that takes seriously the policy invariance ideas from Ng, Harada, & Russell (1999) [another paper I
73.
BORA tackles a core challenge in robot learning: improving vision-language manipulation policies with RL without …
BORA tackles a core challenge in robot learning: improving vision-language manipulation policies with RL without drifting away from the behaviors learned from demonstrations. Instead of letting RL overwrite the pretrained policy, BORA ancho
74.
Does anyone have good studies for what jobs AI has *meaningfully* replaced?
Does anyone have good studies for what jobs AI has *meaningfully* replaced? Looking for a statistical analysis of historical data, present openings and forward looking statements for roles. I know it’s early but still want to read more.
75.
Thanks to (x.com)
Thanks to @Sachin_and_Adam for stopping by to see us. Rock-solid operations are a sum of a bunch of little things, all done well, at scale.
76.
The #ShakeApp is currently going viral with a K-factor of 1.6 for IRL downloads (defined as users who successfull…
The #ShakeApp is currently going viral with a K-factor of 1.6 for IRL downloads (defined as users who successfully shake within 15 minutes of downloading the app).
77.
It's becoming increasingly important to understand the real security properties of cryptographic key storage. A r…
It's becoming increasingly important to understand the real security properties of cryptographic key storage. A related problem is secure provisioning of those cryptographic keys. The shortcut is to just learn from established systems vs.
78.
introducing browser-whisper v1.1 (x.com)
introducing browser-whisper v1.1 • new word-level timestamps support • 9 new models • OPFS cache integration • live mic audio transcription demo • manual download/clear models support • 100% private, offline & open source • new documentat
79.
OpenMed Agent, small Sunday update.
OpenMed Agent, small Sunday update. → queue to stack tasks while the agent is busy → steer to redirect the agent mid-plan → planning view now collapses → theme to swap themes mid-session Quality-of-life wins for builders in the terminal.
80.
Trying a new generative loss: SDMatch. (x.com)
Trying a new generative loss: SDMatch. No discriminator. No denoising chain. Just batch-level distribution matching. Here’s an early CIFAR-10 attempt. Not solved, but interestingly image-like. more info below Thanks @ludocomito for h
81.
A policy that teaches robot hands to touch things the way humans do... not just grab and move, but feel and adjus…
A policy that teaches robot hands to touch things the way humans do... not just grab and move, but feel and adjust in real time. Robot manipulation research often stops at picking up objects and placing them. CGP goes further: it handles
82.
TIP-1034: TIP20 Channel Precompile + TIP-1035: Implicit Approvals
TIP-1034: TIP20 Channel Precompile + TIP-1035: Implicit Approvals We're going to keep baking cost reductions to make payments whether they are on-chain or off-chain feel frictionless We've enshrined the MPP payment channel as a precompile
83.
Workbench now has Slack/Webooks alerts.
Workbench now has Slack/Webooks alerts. Failed jobs → your channel.
84.
Agent Inbox: Every action your AI agent want to take lines up for approval, tagged by risk (x.com)
Agent Inbox: Every action your AI agent want to take lines up for approval, tagged by risk Designed in @paper , animated with @motiondotdev
85.
GitHub 把自己的文档做成了开源工程，远远不止写 markdown。
GitHub 把自己的文档做成了开源工程，远远不止写 markdown。 本质上是一个内容平台：Next.js 全栈 + 自定义渲染管道 + 双仓库自动同步。
86.
Giant migration PRs have never been fun. Something like this would have taken a week of work and a lot of bugging…
Giant migration PRs have never been fun. Something like this would have taken a week of work and a lot of bugging coworkers to review before. Crazy I can knock it out with confidence in a few hours now. Just taking a moment to reflect on
87.
Improved movement. Better hovering. More realistic physics. Now we can race. Who can beat me? (t.co)
Improved movement. Better hovering. More realistic physics. Now we can race. Who can beat me? ↳ https:// flyingcar.evilrabbit.com
88.
how to get better/faster at playing minesweeper: a guide primarily for my nullscape oomfs who wanna get better at…
how to get better/faster at playing minesweeper: a guide primarily for my nullscape oomfs who wanna get better at minesweeper but also for anyone in general
89.
For anybody interested in a tentative dating of the sanskrit literature included at (t.co)
For anybody interested in a tentative dating of the sanskrit literature included at http:// Dharmamitra.org, based on linguistic features and human estimated priors, grouped by categories!
90.
clean. generalized.
clean. generalized. and importantly: fungible / composable