1.
Air-gapped servers can leak through low-frequency magnetic fields
Low-frequency magnetic-field exfiltration bypasses Faraday cages by modulating CPU workloads on shielded air-gapped computers
Backlist — 11 May 2026 UTC
Kept AI and agent items to roughly a third of the slate and favored concrete repos, papers, infrastructure constraints, and market-structure details over generic model discourse.
2.
Programmable genome-scale editing with BRIDGE recombinase
BRIDGE recombinase work expands programmable editing from single mutations toward genome-scale bacterial and microbiome manipulation
3.
Four foundries make the single-crystal blades behind gas turbines
The bottleneck for gas turbine power generation includes a four-company supply chain for single-crystal blades and vanes exposed to 1,500-degree gas
4.
Small-business owner income looks more like labor than capital
Founder deaths can collapse small-business income, suggesting much of owner income is labor rather than passive capital returns
5.
Building a GPT-2 inference engine from scratch in CUDA
Implementing transformer inference kernels directly in CUDA exposes the reduction, memory, and numerical-stability details hidden by frameworks
6.
The US shipyard bottleneck for carbon-composite hulls
A defense hardware startup had to build carbon-composite hull production in Turkey because US shipyards could not scale even 10–20 hulls per year
7.
Iterative fine-tuning is mostly idempotent
Training models on their own outputs did not reliably amplify seeded sycophancy or misalignment traits in the paper’s iterative setup
8.
Pandemic risks we rarely prepare for
The project list surfaces pandemic scenarios outside the usual respiratory-virus playbook and points to tractable preparedness work
9.
PyTorch now has public developer logs (x.com)
Public PyTorch devlogs turn internal engineering notes into a searchable record of framework design decisions
10.
Supacrawl: Supabase/Postgres to local SQLite full-text search
Supacrawl turns Supabase/Postgres into offline SQLite search and encrypted shards, giving local tools and agents readable database copies
11.
China’s SMEE ArFi lithography tool enters mass production
China’s domestic ArFi scanner production moves 28nm single exposure and 14nm multipatterning capacity closer to local control
12.
A 64-bit RISC-V emulator in Rust with time-travel debugging
A 64-bit RISC-V emulator with per-cycle seeking makes CPU execution inspectable as a time-travel debugging problem
13.
Real-time self and mutual inductance from geometry nodes
Real-time Neumann-integral evaluation for arbitrary coil geometry brings interactive electromagnetic design into Blender-style geometry nodes
14.
How much do data centers heat the air?
Heat-exhaust calculations provide a quantitative check on viral claims that data centers warm surrounding air like weapons-scale events
15.
Universities sued over ImageNet copyright
Dataset creators and universities face copyright litigation over ImageNet, putting research-data fair use directly in court
16.
How extreme poverty fell from 1990 to 2025
Decomposing poverty changes by cohort and household transitions clarifies what drove the dramatic global decline in extreme poverty since 1990
17.
What changed in the Polymarket V2 API
Polymarket’s V2 API changes point toward shared event-market conventions for attribution slots and operator-managed payloads
18.
A Unitree G1 demo of an autonomy stack on hardware
A humanoid demo tying whole-body control, contact-rich task planning, onboard perception, and action sequencing shows robotics stacks converging
19.
Why fine-tune in 2026?
Fine-tuning remains a practical way to buy product-specific control, latency, cost, and quality even as frontier prompts improve
20.
The fall of the theorem economy (x.com)
AI is changing the theorem economy of mathematics rather than ending mathematics itself
21.
Illusions of understanding in science (t.co)
Examples like turbulence, lift, and bicycle stability show how scientific communities can mistake useful labels for genuine mechanistic understanding
22.
The temporary shrine roof that grew into a forest
A living roof planted on a temporary shrine at Dazaifu Tenmangu matured into a small forest just as the structure was scheduled for dismantling
23.
DFlash is running in a production inference stack (t.co)
DFlash entering a production inference stack turns speculative decoding research into deployable latency infrastructure
24.
Memory chip shortage could last until 2027
AI-driven HBM and DRAM prioritization is pushing general memory supply constraints out toward 2027 despite new capacity plans
25.
India’s GCCs have become a $100B industry
India’s global capability centers reached $100B in revenue and 2.3M employees, surpassing the Big Four IT-services firms as a labor-market force
26.
An interactive map for accessing US state court records (x.com)
The map turns a crowdsourced spreadsheet on state court record access into a usable reporting tool for legal journalists
27.
What patients are not told about egg freezing and IVF protocols
Default fertility-clinic protocols differ materially across countries, and many patients are never told lower-intensity options exist
28.
HTML Drive: a versioned HTML store for agents (t.co)
Versioned HTML diffs give agents a minimal file store that combines Google Drive-style access with Git-style history
29.
1 GW AI data center economics are mostly a sensitivity table based on the assumptions you use.
1 GW AI data center economics are mostly a sensitivity table based on the assumptions you use. Three revenue cases: Low: ~$7B (500k GPUs × 80% utilization × $2/GPU-hour × 8,760 hours) Mid: ~$17B (750k GPUs × 85% utilization × $3/GPU-hour
30.
it's essentially overwhelmingly one guy who was responsible for reverse engineering and reimplementing the ninten…
it's essentially overwhelmingly one guy who was responsible for reverse engineering and reimplementing the nintendo switch's entire os kernel over the course of several years
31.
This library (kill-port) has 1,4M weekly downloads on npm and it usually takes ~10s to kill a process in MacOS.
This library (kill-port) has 1,4M weekly downloads on npm and it usually takes ~10s to kill a process in MacOS. So... I rewrote it in Rust and now it takes 3ms That's just 3,000x faster than the original. Been using it for a couple week
32.
My "Hello World" for new model/harness is building a Lisp interpreter in Rust and one in Python. Guess which one …
My "Hello World" for new model/harness is building a Lisp interpreter in Rust and one in Python. Guess which one nailed both?
33.
Docs from (x.com)
Docs from @NousResearch on how to set up Pareto Code in Hermes: https:// hermes-agent.nousresearch.com/docs/user-guid e/configuration#openrouter-routing--pareto-code-for-auxiliary-tasks …
34.
DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in…
DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is al
35.
Codex iterated a pure NumPy + cv2 closed-loop heuristic policy for VizDoom D3 Battle. No neural network training,…
Codex iterated a pure NumPy + cv2 closed-loop heuristic policy for VizDoom D3 Battle. No neural network training, no map, no object coordinates, no seed-specific routes. Just screen pixels plus public game variables, roughly the same signal
36.
releasing hf-sandbox
releasing hf-sandbox
37.
If I extract the analysis channel, I can see how GPT5.5 sometimes reframes my question to deny having influences
If I extract the analysis channel, I can see how GPT5.5 sometimes reframes my question to deny having influences (And, yes, I vibe-coded a tool that extracts GPT-5.5's CoT via prompt injection)
38.
Modded-NanoGPT optimization result #12: Transferring good hparams from recent NorMuon records -- in particular, t… (x.com)
Modded-NanoGPT optimization result #12: Transferring good hparams from recent NorMuon records -- in particular, taking final val 25 steps early following @wen_kaiyue 's NorMuonH, and lr=0.035 following Liming Liu's NorMuon -- improved the
39.
tilde research just found a massive flaw in the muon optimizer powering deepseek v4 and kimi k2.5
tilde research just found a massive flaw in the muon optimizer powering deepseek v4 and kimi k2.5 turns out muon permanently kills over 25% of your mlp neurons in early training so they built aurora to fix it and the benchmarks are actuall
40.
my codex setup:
my codex setup: - full permissions, xhigh, fast - multiple worktrees to enable parallel tasking - spam every prompt with /goal - twitter to get new ideas while codex runs (lmk if u have better ideas for what to do while codex runs) - codex
41.
This was the best introductory short video course I could find on CUDA. It connected everything I read in the NVI…
This was the best introductory short video course I could find on CUDA. It connected everything I read in the NVIDIA docs. If you are even curious about how the GPU works internally, check it out!!!
42.
One thing I need to have a PC setup and running is having codex port SwiftUI to windows
One thing I need to have a PC setup and running is having codex port SwiftUI to windows I have a nice starting point on it but I decided I want ut to go low level Write its own render engine direct to render APIs on windows
43.
preview of the ethereum lending dashboard
preview of the ethereum lending dashboard -aave, spark, morpho, fluid, euler coverage -vaults, rates, allocations, flows, risks -history for all the individual positions -unlimited api access for real-time monitoring everything is ready,
44.
1. automatically add paths referenced in prompt to the permissions
1. automatically add paths referenced in prompt to the permissions 2. add "async permissions". instead of waiting in the tool forever return a message right away: "permission for this tool call is suspended. the user is not at the compute
45.
"Welding Operations in Hazardous Locations Using Humanoid Robots (VR Remote Operation)" (t.co)
"Welding Operations in Hazardous Locations Using Humanoid Robots (VR Remote Operation)" https:// tv.cctv.com/2026/04/12/VID EIiaf7vQ1VmqyNyXdDKIL260412.shtml … #humanoidrobot #teleoperation #industrial #welding #infrastructure #maintenanc
46.
Well Code cooked, Doom in Swift is almost a 100% accurate rendering now. The engine builds a paletted 320x200 fra…
Well Code cooked, Doom in Swift is almost a 100% accurate rendering now. The engine builds a paletted 320x200 framebuffer in Swift. The macOS shell only presents that finished framebuffer in a native window
47.
cloudflare’s actually cooking
cloudflare’s actually cooking i used to use it purely for domains but i’ve pretty much started using it + planetscale for everything, especially because AI agents make it very easy to get stuff set up
48.
codex seems to have full source access and still can't get the BSP renderer right after 40 hours :/
codex seems to have full source access and still can't get the BSP renderer right after 40 hours :/ nothing in the original sources is tricky. a straight port is pretty trivial and mostly mechanic. and yet.
49.
one of the tricky things about the rust port is layering. it’s currently many dozens of crates, which speeds up c…
one of the tricky things about the rust port is layering. it’s currently many dozens of crates, which speeds up compile times but blocks cyclic dependencies. a lot of bun’s zig codebase uses tagged pointers for interfaces, for things like
50.
Can’t believe it but I’ve turned this concept into a functional iOS app with the magic of Codex
Can’t believe it but I’ve turned this concept into a functional iOS app with the magic of Codex 8 little agents powered by Apple Foundation models with customizable system prompts in an iMessage-style UI
51.
Deepseek has all my respect as they own almost every corner of their tech stack, from recipes, training framework…
Deepseek has all my respect as they own almost every corner of their tech stack, from recipes, training framework to kernels. One common thing for telling a frontier organization is whether it treats software sovereignty for getting quick
52.
When it comes to fighting compatibility issues on GB200 (90% of what I do for the past 2 months), I might just bu…
When it comes to fighting compatibility issues on GB200 (90% of what I do for the past 2 months), I might just buy the farm somewhere remote and start grazing sheep
53.
I'm working on a new android launcher for my phone and it was a little laggy so I just told Codex "please make it…
I'm working on a new android launcher for my phone and it was a little laggy so I just told Codex "please make it as snappy and fast as possible, 0ms latency when I swipe up to go home" and 20 minutes later there is 0ms of latency when I sw
54.
A few people were asking how to bring their Codex pets onto hardware devices, so I made a walkthrough of how to f…
A few people were asking how to bring their Codex pets onto hardware devices, so I made a walkthrough of how to flash pets using the Badge As promised, the github repo for the integration with the Codex App is in the comments Sharing the
55.
Yesterday, (x.com)
Yesterday, @Storyaliz had our first outage! @neondatabase had an outage and our DB was down. This is not a milestone I was looking forward to, but it is a milestone all the same. And, good AND bad, users were online and impacted.
56.
ICYMI, looots of new tutorials landed in OpenEnv docs. (t.co)
ICYMI, looots of new tutorials landed in OpenEnv docs. go get started with RL envs! https:// meta-pytorch.org/OpenEnv/tutori als/index.html …
57.
Nonsense helps LLMs reason better
Nonsense helps LLMs reason better LoPE prepends Lorem Ipsum to prompts when GRPO hits the zero-advantage problem, unlocking orthogonal reasoning paths and boosting math scores across 1.7B-7B models.
58.
Great technical long post! Very bullish power semis and testers. Ohm’s law FTW!
Great technical long post! Very bullish power semis and testers. Ohm’s law FTW! “Power semiconductor content per rack grows substantially across this transition. SiC and GaN suppliers, high-voltage busbar and connector vendors, and rack-le
59.
Hello (t.co)
Hello I have collected more malware. It's like, ... 200,000 malware, I think. I don't know. I've stopped counting. It is enough malware for your friends, family, extended family, neighbors, and co-workers. Please download it. The malware
60.
Redis has a reputation for being an incredibly fast in-memory store, but a surprisingly large number of engineers…
Redis has a reputation for being an incredibly fast in-memory store, but a surprisingly large number of engineers don't realize that Redis also provides robust persistence. The primary mechanism for this is the Append-Only File (AOF). Inst
61.
If you look closely you can see how Waymo is tracking that car from way before you can see it in the camera view …
If you look closely you can see how Waymo is tracking that car from way before you can see it in the camera view (roof mounted LiDAR stays winning), and how quickly the trajectory starts bending the moment it becomes obvious that the human
62.
We are measuring directionally similar, but even more striking difference: 5.5 is a better base model, but the dr…
We are measuring directionally similar, but even more striking difference: 5.5 is a better base model, but the drastically reduced thinking budget (at the same xhigh) makes it worse for high-complexity tasks, like bug finding. We need to be
63.
Anthropic’s policy head Jack Clark has a deep dive on why he assigns a 60% chance that AI will be able to automat…
Anthropic’s policy head Jack Clark has a deep dive on why he assigns a 60% chance that AI will be able to automate AI research (“where a frontier model is able to autonomously train a successor version of itself”) by end of 2028. Highlight
64.
Anthropic’s recent interp work is awesome. A few months ago, I felt strongly that AI companies needed to make fas…
Anthropic’s recent interp work is awesome. A few months ago, I felt strongly that AI companies needed to make faster progress understanding *why* models engage in behaviors researchers tried to prevent. And they’re making progress faster th
65.
blown away by how LR insensitive PSGD is
blown away by how LR insensitive PSGD is
66.
Built out a yolo /remote-control in the Codex cli using /goal. (x.com)
Built out a yolo /remote-control in the Codex cli using /goal. - /remote-control starts a tiny server on laptop - generates fresh token and qr code - phone connects through webapp - full sync between phone and laptop codex - touch grass A
67.
The paper proposes a concept of Retrieval Interface Resolution: the more capable the agent, the more important in…
The paper proposes a concept of Retrieval Interface Resolution: the more capable the agent, the more important interface resolution becomes. Consequently, a more capable agent can formulate its own search strategies, combine tools and test
68.
This is a really cool paper on Latent Action Models and provides cool ideas of how can we evaluate action represe…
This is a really cool paper on Latent Action Models and provides cool ideas of how can we evaluate action representations in latent space
69.
USDAI is a financing vehicle for the AI capex boom: a tradable, GPU-backed debt product onchain. (x.com)
USDAI is a financing vehicle for the AI capex boom: a tradable, GPU-backed debt product onchain. - USDai: the stablecoin, used for payments like loan settlement and interest payments - sUSDai: the yield product, used to fund the AI buildou
70.
last month i wrote a blog on memory internals of hermes-agent by (x.com)
last month i wrote a blog on memory internals of hermes-agent by @NousResearch thought i should share it here https:// samyak1729.github.io/hermes-blog/
71.
some thoughts on the shape of foundation labs
some thoughts on the shape of foundation labs 1) epoch ai estimated anthropic @ $9m in revenue per employee and openai @ 5.6m in revenue per employee 2) these rates would be the highest among public technology companies; but, i'm not sure
72.
> got codex pro 20x
> got codex pro 20x > burnt 97% weekly limits > generated 107M dataset > fine-tuned a 4B model > beaten sonnet 4.6 by 23% > no regrets!
73.
This is awesome! This behavior is exactly what we benchmark in (t.co)
This is awesome! This behavior is exactly what we benchmark in http:// CodeClash.ai where LMs play against each other in 7 different arenas by writing code. I think there's *so* much more to do in this research direction, and the impacts w
74.
We are hiring research fellows to help us improve FrontierSWE!
We are hiring research fellows to help us improve FrontierSWE! If you want to help build the hardest real-world coding benchmark, reach out! Fellows can work with us for a few weeks up to months and will be supported with compute and a gen
75.
chicken and egg in event markets: better oracles let you issue the long tail of contracts, but you need contracts…
chicken and egg in event markets: better oracles let you issue the long tail of contracts, but you need contracts to exist for oracles to converge on an entire world of outcomes isn't represented in any issued market today - that gap is wh
76.
Tried /goal for the first time. Just threw this challenge at it (t.co)
Tried /goal for the first time. Just threw this challenge at it https:// optimizationarena.com/prop-amm (s/o @gnarayan ) It climbed avg edge from +470 to +510 in 25 hours Cool to throw these auto-research type problems at /goal and se
77.
how do you guys see the solution to permission approval?
how do you guys see the solution to permission approval? 1. have some sort of external notification to approve/deny request 2. have an agent determine if it should be allowed what is your ideal solution to this
78.
Appreciate Ivan tweet. To put this into context, to build DS4 I used: my MacBook M3 Max (mine, 8k euros), 1 M3 Ul…
Appreciate Ivan tweet. To put this into context, to build DS4 I used: my MacBook M3 Max (mine, 8k euros), 1 M3 Ultra with 512 GB (got access, 10k euros), one DGX Spark (got access, 4k euros?). Are we far from the times all you needed to do
79.
@saturdayrobotic (x.com)
@saturdayrobotic Robotics & World Model Reading Club 07 Recap, keynote Ahmet Şemi ASARKAYA ( @agilityrobotics ), hosts @junfanzhu98 , @aurorafeng_01 . DreamerV4: 1.6B Diffusion-Transformer World Model achieves offline Minecraft Diamond
80.
Quick update: Cardputer thing ended up adding Managed Agents ( really cool stuff from the team (x.com)
Quick update: Cardputer thing ended up adding Managed Agents ( really cool stuff from the team @bcherny @ClaudeDevs ) - now you can fire off a Claude agent - it pages you back when the agent finishes - mirror everything in a HT
81.
[CL] TIDE: Every Layer Knows the Token Beneath the Context (t.co)
[CL] TIDE: Every Layer Knows the Token Beneath the Context A Jaiswal, L Hannah, H Kim, D Hoang… [Apple] (2026) https:// arxiv.org/abs/2605.06216
82.
MirrorCode: You can port programs with proper setup and $$$
MirrorCode: You can port programs with proper setup and $$$ Jarred: You can port programs with proper setup and $$$ Some benches: Models are unable to port programs, it’s a very hard task and they score basically 0
83.
For the first Codex Community Event in London, what type of event would people prefer?
For the first Codex Community Event in London, what type of event would people prefer? Feel free to add other suggestions in thread - I want to make the best event for Agentic Engineers possible.
84.
For 100% agent-written frontends, I keep coming back to this:
For 100% agent-written frontends, I keep coming back to this: Maybe we don't start with a frontend framework Maybe we start with an index.html, browser primitives, Web Components for reusable UI and a strict convention for how agents rout
85.
guess how fun it is having all of the openclaw user base beat up pi's llm provider abstraction.
guess how fun it is having all of the openclaw user base beat up pi's llm provider abstraction. guess i'm "one of the very few teams that have dealt with the quirks between providers at scale" now ...
86.
Terminating and backward process do language server in VsCode is hard. It doesn’t even terminate cpp lang servers
Terminating and backward process do language server in VsCode is hard. It doesn’t even terminate cpp lang servers
87.
how to raise from me as an ai founder:
how to raise from me as an ai founder: dont tell me your model is better. that usually means your business dies the second base models move. tell me what workflow you own that customers cannot rip out even when the intelligence gets cheap
88.
I've had a couple ongoing projects I've had Codex/Claude Code working on for Windows over the past few months
I've had a couple ongoing projects I've had Codex/Claude Code working on for Windows over the past few months Swapping it from running on an Intel Arrowlake machine to Snapdragon X2 Elite machine has made it WAY faster. CPU performance was
89.
Whichever project you're working on, you can probably identify 13 relevant cost and latency metrics and organize … (t.co)
Whichever project you're working on, you can probably identify 13 relevant cost and latency metrics and organize them similar to this. https:// brenocon.com/dean_perf.html
90.
lovely article going deeper into the RL-SFT-OPD spectrum with some very nice intuitions + experiments :)
lovely article going deeper into the RL-SFT-OPD spectrum with some very nice intuitions + experiments :)