Top 90 curated tweets ranked for substance on 16 May 2026 UTC.
The release bundles 367 commits from 202 contributors and targets current bottlenecks in reasoning-model serving, Blackwell MLA kernels, and distributed KV caches
The largest US grid averaged $136.53/MWh in Q1 as AI load growth became a directly measurable electricity-price shock
A GPU programmer walks through six kernels and profiling bottlenecks showing that Python-hosted CuTe DSL can approach vendor-library performance on Ada Lovelace
Event cameras reveal a bullet’s passage through a target by its shadow, showing a sensing mode ordinary frame cameras miss
The Redis Object abstraction explains how strings, counters, and other key values get compact internal representations while preserving a simple API
Automatically generated open-ended tasks trained long-horizon coding agents better than expert-curated data, shifting the bottleneck from curation to verification
A diagnostic dialogue system that tracks patient state and accepts uploaded images improved clinical history-taking in a setting closer to real patient encounters
A trading firm with 3,500 employees and $16B in quarterly trading revenue is operating infrastructure that looks more like a frontier AI company than a bank
Activation and gradient spectra can expose representation geometry, forecast token efficiency early, and separate real learning gains from throughput gains
A teardown of an Infinera photonic integrated circuit connects chip structures to dual-polarization 64QAM optical networking at 96 GBaud
Columbia’s 125th Street station accessibility project illustrates how real-world infrastructure costs can be dominated by process, coordination, and construction overhead
HNSW’s layered graph search makes nearest-neighbor retrieval fast by borrowing the same intuition that lets skip lists avoid scanning every element
The release handles difficult Unicode cases such as Hangul composition, Japanese dakuten, and emoji without drifting from the reference tokenizer behavior
Gen-El-Mec milled a titanium part with 0.014-inch walls and ±0.0005-inch tolerance, a useful reminder of how extreme conventional machining can get
$3.4B in crypto was reportedly stolen in 2025, with more than 80% of first-half losses traced to hot wallets, compromised laptops, poisoned packages, or phished signing infrastructure
Kalshi has flagged more than 400 suspicious trades this year, already more than double its 2025 total, while Polymarket has seen a similar rise
A simple prepublication approval flow would address AI-slop and fake-author problems without relying on shared paper passwords or retroactive punishments
A magistrate judge in eastern Oklahoma issued an opinion with extensive machine-written text, pushing generative AI from law-office drafting into judicial output
Early research suggests many expensive techniques built to attract human attention and influence choice fail when the chooser is an autonomous agent
A Bloomberg-reported investigation into BlackRock private credit follows off-cycle asset markdown disclosures and raises the question of how illiquid credit is being priced
A five-hour session yielded three hours of robot data, with frozen arms and unobstructed motion making VR teleop feel more practical than leader-arm setups
Reimplementing ULTRA-style retargeting shows reinforcement learning can preserve human-object interaction and plausible motion better than naive kinematic retargeting
A sub-$10k in-house machine was enough to start manufacturing and sell the first million dollars of product, showing how scrappy tooling can beat waiting for capex
Page tables, page faults, mmap, page cache, swapping, and NUMA explain many of the design choices behind databases and data-intensive systems
The Tillman Water Reclamation Plant turned unpopular civic infrastructure into a public landmark good enough to stand in for Starfleet Academy
Agent configuration files now depend on Claude Code’s permissive parser behavior, forcing other tools to decide whether to reproduce a broken YAML dialect
Continually consolidating past experiences into compact memories can make agents worse, so self-improvement through memory is more fragile than it appears
A small aerospace shop compressed CAD-to-part turnaround for aluminum drone hardware into a single afternoon, showing the new tempo of hard-tech prototyping
DeepSeek V4 is now merged into SGLang main with v0.5.12. What we shipped at launch: ShadowRadix: native prefix caching for V4's hybrid attention HiSparse: CPU-extended KV for sparse attention (up to 3× long-context throughput) MTP spec
Just now hitting me that Bun is the runtime for (likely) 10B+ in revenue for Anthropic: Claude Code, Cowork, Claude Agent SDK, and they re-wrote it in a new language in 6 days. Insane level of conviction in recursive self-improvement
Open-ended training data scaling via AI is the moment the data bottleneck for coding tasks effectively disappears. The new bottleneck becomes eval quality: if AI writes the tasks and the tests, who grades the grader?
the "Mythos 69%" headline is misleading GPT-5.5 Codex sits at 41%, looks close. on T1 (full control, the only tier that matters): Mythos: 16 envs GPT-5.5 Codex: 2 Opus 4.7: 0 9x gap, not 1.7x. baseline Mythos (no hints) hits MORE T1s th
I don’t really want to multithread across 10–20 minute Codex requests. I want a realtime conversation where I dump context once, then hand the goal off to a long-running background agent that periodically updates me on progress.
Microsoft just released GridSFM on Hugging Face A graph foundation model for power grid optimization that learns across 54 topologies to predict voltages, dispatch, and feasibility in one shot.
i mitm'ed vscode and it appears nothing is actually send if your telemetry setting is set to off. the codepaths are still executed. which is why the payloads above got successfully logged to disk after instrumentation.
LMFAO. Opus 4.7 just wrote me a fix for a bug, wrote tests, wired it all, it all passed. Great. then it goes - same session; "now let me revert all my changes as a sanity check and verify that the tests no longer pass" BRO WHAT.
Note that theyre actually using block scaled fp8 x fp4 blackwell tensor cores.
I've released 4 quite distinct evaluations in my PhD now: code falsification, long-horizon execution, research plan generation, and now forecasting agents. GPT models have been far ahead everytime. The wide range of tasks OpenAI post-train
What a beautiful sight @CloudflareDev finally solved the $25/day durable object cost issue a 3-second setInterval logic in YJS awareness that was stopping the durable object from hibernating cons of building your own real-time multip
Docker in docker is not enough for agent sandboxes, agents need VMs. Nested Virtualization, BPF, real networking, RootFS and fast IO all matter for actually useful agents.
How do we train LLM agents to be better continual learners that self-evolve after deployment? Introducing SkillOS. We propose an RL recipe for training skill curation that distills wisdom from past experiences via indirect and delayed fe
tldr: Anthropic needs more GPUs. OpenAI needs better quality demand. xAI had supply and no demand. xAI figured it was easier to generate developer demand through the Cursor partnership/acquisition and monetize extra capacity by leasing som
200 Gb InfiniBand NIC with custom cooling for my GPU cluster
Big inference day with @tuhinone from @baseten at MS&E 435 yesterday (and the Cerebras IPO that AM) We discussed compute scarcity amid the rising inference demand, model routing, & open source! Video soon.
We're now on Day 4 of nonstop autonomous operations with F.03 humanoid robots running 24/7 until failure
Since I'm between jobs, I've been having a lot of fun vibe-coding with public tooling. First drop: a clean PyTorch impl of the Gradient Moment metric from our recent paper (arXiv:2603.20155). https:// github.com/ehoogeboom/gra dient-mome
New article: a visual tour of recent LLM architecture advances, from Gemma 4 to DeepSeek V4. I focus on long-context efficiency tweaks like KV sharing, per-layer embeddings, layer-wise attention budgets, compressed attention, and mHC. Lin
DSA or HISA have their selection baked into their attention kernel and, wiring a new kernel for every gpu arch is a cumbersome process, we avoid this and make selection out of attention. We have a selection kernel which does bitonic topk m
DEI = datacenter, electricity, and infrastructure
We found and fixed two issues that could explain this degradation of the capability of GPT-5.5 in Codex over the last ~ 48 hours. We are monitoring over the coming hours to fully confirm and I will reset usage limits this evening. Apolog
Oxford researcher did a full deep dive on China’s “proxy station economy” for Claude/OpenAI access - https:// chinatalk.media/p/how-to-buy-c heap-claude-tokens-in … The cheap pricing is basically just arbitrage. People reverse engineered
New paper: LLMs are now used for high-stakes real-world decisions, but can their numerical predictions and uncertainty estimates be trusted? We built QuantSightBench, a benchmark to measure how well frontier models forecast numerical outco
There's a massive data gap holding back AI coding agents: millions of LeetCode problems are there, but only hundreds of open-ended tasks. FrontierSmith fixes that by automatically converting closed problems into open-ended ones — and the r
Fun visit to the @DeltaInstitutes podcast where we discussed all things agent improvement ranging from data to evals to post-training! Sleeper show. Honored to be featured amongst some technical beasts
actually maybe it's just time to design new challenges: *moe (!) *intently multilingual *documented/releasable data (and mode failure) *at least partial synth with reasoning traces throughout (SYNTH/Zyphra) *benchmark targets on top of loss
David Silver RL Course (Lecture 8): Integrating Learning and Planning AlphaGo is a beautiful example of integrating learning and planning. The policy network gave AlphaGo intuition: which moves look promising? The value network gave Al
FLX-5024 256 × 8 SRAM Texas Elektronikgruppe, Inc.
Can someone explain to me how to get an LLM to perform in a very high-quality, high-token way, like @RefineDotInk ? When I build a skill and run it using Codex or Claude it does a mediocre job. What am I missing? 1/
How do we teach LLMs to learn what they can forget ? A technical break down of the Self-Pruned KV Attention mechanism with fun ML techniques The SP-KV mechanism works by projecting a token's hidden state into a sigmoid to get a scalar "fu
Scouting places to build the next AI data center
Qwen-Image-VAE-2.0 from Alibaba achieves SOTA VAE reconstruction Training on billions of images with Global Skip Connections and asymmetric architecture, introducing OmniDoc-TokenBench for rigorous text-rich document evaluation at high com
getting few DMs on how/if this works: yup, one ‘import’ command *literally* creates new accounts or links to all the providers in the stack, provisions them in batch, and returns the full set of keys & env vars you need. wild scenes
Multimodal AMIE now published @NatureMedicine ! This is from past work @GoogleDeepMind where we studied patients uploading images during diagnostic dialogue. We found that a multimodal reasoning harness that tracks a patient’s state gre
my number 1 agent tip is "make the agent prove it understands things". everything boils down to that. the UVs on this book are the best example I have. if you don't do this and then ask the model to do something like "put some text on the
Warden Pi branch is still optimizing itself 24 hours later with little success. Doing this against evals aint no fun yall...
The Opus 4.7 insistence on `| tail -30` or `| head` at the end of commands, like some constantly confused model from last year, is enough of a regression on its own to push me back to 4.6.
You don’t give your agents root? Every other day I have agents running tcpdump or strace or some other task that needs root and needs to be inside its dev loop or they become horribly inefficient.
I have just tested the released VGGT-Omega checkpoint from the VGGT team. TL;DR: It is an absolute beast for outdoor scenes or scenes with larger depth ranges. Here are some results (1/10)
A new way to pre-train language models that gives quite faster training. Normal causal LLMs predict the next token t+1 But this method predicts a bag of future tokens: (t+1, t+2, t+3, …) in a single step using token superposition. Instead
複数AIエージェントに強化学習（RL）をかける方法を体系化した論文が出た（https://arxiv[.]org/html/2605.02801v1）。 単一AIのRLをそのまま複数AIに使うと失敗する。理由が面白い。 最大の原因は「クレジット拡散」（credit
Auto Research for compilers
imagine being a gen1 pokemon AI, losing to the same strategy that every kid has used since 1996, only this time it's an LLM trained with a hundred jiggawatt hours of electricity that thinks a quadrillion times faster than you, moves more da
. @bulktrade solves this with structural priority queues. Sequenced within native solana validators by order type and fifo 1. Cancels 2. Maker 3. Market, limit, ioc 7ms+ time gap between maker and taker
this is one of the Sydney screenshots that Sonnet 4.5 was referring to (click to expand)
Honestly 5.5 is being so stupid about this heap groom that I’m tempted to think it’s sandbagging. It keeps picking the *least* sensitive stuff to overwrite, like its own data?
I'm joining @GraySwanAI 's R&D team this Summer to work on AI security for frontier models. Excited learn from the best in LLM attacks and defenses!
Actually kind of surprised that I seem to be the only person on an AI coding mobile app when out and about in SF Had to build a lot of infrastructure to be able to connect to real devices in our lab via web on my phone… but now
Consolidated one of my workflows today and it ended up creating a clean map of how institutional-side prediction market structuring + execution currently looks actor and flow wise. Sharing it as someone might find it useful. Sell-side t
Ok this is pretty cool > http:// zero.coey.dev > http:// github.com/acoyfellow/zer o-cloudflare-hello …
Thrilled to see those promising numbers! Same finding on our end with NanoRollout: cross-scaffold generalization basically doesn't happen out of the box -- something the field should be talking about more.
Chronic disengagement location in Amsterdam for FSD In the past few weeks I found this spot in Amsterdam where FSD makes the same mistake, repeatedly It doesn’t really matter if you come from a sidestreet like this or if you’d be coming
you can turn telemetry off by ... disabling the GH Copilot extension. No settings anywhere. seems to add to auto-complete lag (and i'm not talking about LLM suggestions or similar, just the plain old intellisense path next to the LLM sugges
Has anyone built a multi-user agent on a platform like GitHub, where you might need auth, but you cant disclose private connection links?
LLM APIs need to return cost information in their response alongside tokens literally everyone is using models[dot]dev data to approximate this - we see so many reqs to its api but this is just sticker pricing, won't reflect discounts, et
Grok CLI has great support for Plugins and Skills. Installing the Vercel Plugin gives Grok cloud deployment superpowers. Watch this creative coding website be generated with Grok and hosted seamlessly on Vercel ↓ http:// vgrok.vercel.app
This is interesting just on the merits and also “This is a segment that makes no sense to any broadcaster on any economical model in history of television, and yet makes excellent sense for all participants here.”
How to be good at ai coding, the parts orthogonal to the llm / ai skills and intuition
just realised u can have agents do best-of-n by deploying multiple subagents with the exact same prompt...
We evolve user simulators so realistic, in live human evals people can’t tell the difference between them and real people! Whereas the baseline user simulator is identified as a bot ~half the time. If you need to train AI agents with more
My goal with MachinenVM was to learn to build a microvm. I understand maybe 33% of what it's doing, but my goal has always been to understand the machine from first principles: Did http:// LinuxFromScratch.org as a teenager, and Nand2Tetri