1.
OpenAI says the TanStack npm attack nearly reached signed official software
A compromised npm dependency chain came close to putting attacker-controlled code into trusted OpenAI releases
Backlist — 15 May 2026 UTC
Top 90 curated tweets ranked for substance on 15 May 2026 UTC.
2.
Microsoft Edge Enhanced Security Mode contained 23 RCEs (x.com)
A feature designed to harden risky browsing exposed a large attack surface inside Microsoft’s WebAssembly interpreter
3.
FutureSim: replaying the web day by day for continual-learning evals
Agents can be tested on whether they update forecasts as real events unfold rather than on static benchmark snapshots
4.
Cross-request LoRA batching makes SLM inference 44x faster
Packing tokens from many adapters into one batched pass turns small-model serving from adapter-per-forward overhead into efficient shared inference
5.
Running RAG in production is not a demo problem
Many RAG failures come from retrieval, ranking, freshness, observability, and data plumbing rather than from the language model itself
6.
Snapshot and resume 32GB/60GB sandboxes in 1–2 seconds
Fast snapshotting large dev environments changes the economics of short-lived agent sandboxes and heavyweight compile/test loops
7.
Cerebras IPO returned billions to early VCs
Benchmark, Foundation, and Eclipse turned early Cerebras checks into multibillion-dollar stakes, showing how one hardware IPO can return entire funds
8.
What is going on with H200 chip sales to China?
US and Chinese internal politics have produced a confusing situation where China may formally accept chips its AI labs already want
9.
About 4% of US GDP depends on rare earths
Industries worth roughly $1.2T of US output rely on rare earth inputs that remain vulnerable to Chinese supply disruption
10.
Let communities decide on data centers if developers cover grid costs
A clean data-center policy separates local disruption decisions from the requirement that developers pay for generation and grid upgrades
11.
Cloudflare R2 SQL gets joins, subqueries, and multi-table queries (x.com)
SQL semantics on Cloudflare’s object-storage-adjacent infrastructure now support more realistic relational workloads at the edge
12.
Investigating Fluid’s $21M bad-debt cleanup
An $8M piece of Fluid’s depeg remediation appears to have come from an uncollateralized credit line rather than from the treasury as presented
13.
Tag hallucinated references instead of banning arXiv authors?
Hallucinated citations can be flagged with public evidence without turning a preprint server into a career-damaging gatekeeper
14.
538’s data archives were destroyed
Deleting decades of election and polling data erases a public research archive for no technical or scientific benefit
15.
Why Keytruda required decades and $46B of R&D
Keytruda’s success depended on a century of immuno-oncology work, years of skepticism, thousands of trial patients, and massive development spend
16.
GS-Playground: Gaussian splats plus physics for 104 FPS robot simulation
Photorealistic 3D Gaussian Splatting combined with a fast physics engine gives robots vision-rich training scenes at over 100 FPS
17.
Seven robotics wedges that are not humanoids
Actuators, magnets, and industrial subsystems may be better robotics entry points than building another humanoid platform
18.
Meta’s KV self-pruning uses 15–35% of memory at parity (x.com)
Letting an LLM learn which KV pairs to forget cuts memory use sharply while preserving performance and improving throughput
19.
Why critic-free RL works in LLM post-training
Sequence-level rewards can assign hidden token credit because gradients from positive and negative rollouts cancel in structured ways
20.
Lighthouse Attention: sparse pretraining with dense inference
Sparse attention can accelerate pretraining while still producing a model that works with dense attention at inference time
21.
OxCaml and the case for garbage collection in hot/cold paths
Most code paths benefit from GC, while the small performance-critical slice can be engineered to avoid allocation pressure
22.
SpacetimeDB’s architecture has a real performance argument
Pushing entire transactions into the database can produce very high QPS when metadata scaling is the bottleneck
23.
Will compute become an actively traded capacity market?
GPU capacity may evolve into either bilateral long-term contracts or a liquid commodity market, with very different consequences for inference providers
24.
ByteDance opens Cola DLM (t.co)
A hierarchical latent diffusion language model for text is now available with weights and code for outside inspection and follow-up work
25.
How Instagram Close Friends is actually used (x.com)
A 500-person survey and 20 interviews show how a seemingly simple privacy feature creates social meaning and confusion
26.
Compare US city boundaries by dragging them around a map (t.co)
Moving and comparing city outlines makes the weirdness of American municipal boundaries visible in a way static maps do not
27.
Floodify ran 65,000 fake social accounts for clients
A single attention-for-hire company operating tens of thousands of dummy accounts illustrates how much online popularity is manufactured
28.
Hyperliquid’s pre-IPO CBRS market and public historical data (x.com)
Pre-IPO perpetuals let an onchain venue participate in price discovery while exposing trades, candles, order books, and positioning data for verification
29.
An acoustic resonator simulation self-organized into a toy universe
A chaotic resonator simulation stabilized into solitons, loops, and knots without being explicitly designed to do so
30.
I interviewed (x.com)
I interviewed @bubbleboi about his ratings of AI supply chain bottlenecks. We talked about DRAM, advanced packaging, CPO, HBF, PCBs, power delivery, etc. 0:00 HBM, DRAM, the cartel 7:24 Silicon photonics, CPO, Lumentum lasers 11:35 Adv
31.
Step 4 to achieve truly serverless GPUs for AI inference: skip over unserializable inference engine setup steps l…
Step 4 to achieve truly serverless GPUs for AI inference: skip over unserializable inference engine setup steps like CUDA graph capture and Torch compilation by stacking GPU snapshots and CPU snapshots.
32.
The ships keep coming. Wrapping up my first week with some inference optimizations that reduce p50 latency by 50%…
The ships keep coming. Wrapping up my first week with some inference optimizations that reduce p50 latency by 50%. The effect is bigger on long tail latency - up to 80% faster on very long docs
33.
Introducing FutureSim: where we replay a temporal slice of the web and let agents forecast real-world events over…
Introducing FutureSim: where we replay a temporal slice of the web and let agents forecast real-world events over time FutureSim replays the web day by day. Agents start on Jan 1, 2026 (past their knowledge cutoffs) with date-gated access
34.
thrilled to introduce my and (x.com)
thrilled to introduce my and @timhwang 's new latent social archetype generalization and narrative alignment eval more cooking
35.
pytorch profiler kind of day
pytorch profiler kind of day nsight compute kind of day
36.
Wondering what strategies the harness adopt to achieve this. (t.co)
Wondering what strategies the harness adopt to achieve this. Our BenchJack achieves 100% on top benchmarks — with reward hacking! Check it out https:// github.com/benchjack/benc hjack …
37.
"cloudbox" (t.co)
"cloudbox" Fresh Cloudflare computers for agent repo work - clone a repo, run commands, verify, return one artifact - every step recorded as a receipt you can audit - runs in a Cloudflare Container, deploys to your account http:// cloud
38.
Beau Rothrock had been at (x.com)
Beau Rothrock had been at @AngelList for two months when he walked into a Redshift-to-Snowflake migration in deep trouble, already two months behind schedule. He had a 5-week window to migrate all 14,000 dashboards and reports AngelList
39.
Open-ended coding training data may no longer be the bottleneck: AI can scale open-ended tasks—and even outperfor…
Open-ended coding training data may no longer be the bottleneck: AI can scale open-ended tasks—and even outperform human-expert curation. FrontierCS team is releasing FrontierSmith: a system for synthesizing open-ended coding problems at s
40.
"If TPU v9 upgrades the topology, optical module speed, and port ratio at the same time, a roughly 4x increase in…
"If TPU v9 upgrades the topology, optical module speed, and port ratio at the same time, a roughly 4x increase in ICI bandwidth versus TPU v8 may not be entirely out of reach. This is likely not just a matter of "buying more optical modules
41.
Check our new work on explaining critic-free RL of LLM (x.com)
Check our new work on explaining critic-free RL of LLM TLDR: Advantage sign is a poor predictor of token updates. Our Cancellation Hypothesis: grads from +/- rollouts cancel out, inducing hidden token credit assignment. Huge thanks to my g
42.
im splitting my day between writing cute dsl kernels and learning sglang architecture. this has been the most fun…
im splitting my day between writing cute dsl kernels and learning sglang architecture. this has been the most fun split ive had in a while.
43.
This is like a good stress test for optimizers.
This is like a good stress test for optimizers. Kaon is basically Muon/lmo + spectral noise. It preserves the singular vectors of the gradient and randomizes only the positive singular weights. For exchangeable noise, the conditional expec
44.
Anthropic CFO Krishna Rao on Mythos:
Anthropic CFO Krishna Rao on Mythos: "We had an open-source code base that a prior model found 22 security vulnerabilities in, and Mythos then found 250. That is kind of scary, but that informed the way in which we released it."
45.
looking at data is underappreciated! In the main nanogpt speedrun track we filter to at most the first 2048 token… (x.com)
looking at data is underappreciated! In the main nanogpt speedrun track we filter to at most the first 2048 tokens per document, to prevent a single gradient update from getting dominated by a single OOD document. Each step gets at least 64
46.
The latest OpenClaw release is ~3.5x faster
The latest OpenClaw release is ~3.5x faster We run end-to-end RTT tests against every published npm release, every 6 hours, over real message channels (here: Telegram, using the brand new bot-to-bot communication). No more silent regressi
47.
Why do you need to talk about CUDA streams and CUDA events for a blog post on Continuous Batching?
Why do you need to talk about CUDA streams and CUDA events for a blog post on Continuous Batching? I have recently started reading more about LLM inference optimization. Upon asking around, I was quickly greeted by the term "Continuos Batc
48.
I'm hiring for 2 roles in Asta/AI for Science (x.com)
I'm hiring for 2 roles in Asta/AI for Science @allen_ai Research Engineer: RL/post-training for hypothesis generation, long-horizon agents, continual learning PhD Research Intern, Fall 26: Designing new rewards beyond surprise & novelt
49.
Codex team:
Codex team: You already have the beautiful right sidebar with git, files, and the browser Why add this floating card outside of the functional right sidebar? This could be a tab just like browser/files are It's just going to make growing
50.
made something like this! it even supports websocket and HMR so you can run a full dev server. starts much faster…
made something like this! it even supports websocket and HMR so you can run a full dev server. starts much faster than Cloudflare tunnels and it lets you use a stable domain without logging in
51.
I just realised you can use capnweb to build a lightning fast poor-man's version of cloudflare tunnels
I just realised you can use capnweb to build a lightning fast poor-man's version of cloudflare tunnels You just need to write a tiny durable object class that hosts a capnweb session, and then write a tiny client side utility that connect
52.
ByteDance Seed just released Cola DLM on Hugging Face
ByteDance Seed just released Cola DLM on Hugging Face A hierarchical latent diffusion model for text that separates global semantics from token generation. It beats AR and LLaDA on 8 benchmarks with 2B params.
53.
1/
1/ Millions of parameters dedicated to LLM safety fine tuning, and it turns out the entire guardrail can be dismantled by flipping exactly one neuron.
54.
Excited to see MiniMax in action inside open-multi-agent! It automatically breaks goals into DAG tasks & runs th…
Excited to see MiniMax in action inside open-multi-agent! It automatically breaks goals into DAG tasks & runs them in parallel
55.
I get a lot of questions about why Kubernetes isn’t the right foundation for sandbox infrastructure.
I get a lot of questions about why Kubernetes isn’t the right foundation for sandbox infrastructure. It was built for stateless micro-services with predictable traffic patterns, and databases that run forever. Sandboxes are different, the
56.
Amp's `smart` mode uses Opus 4.7 from multiple redundant providers (Anthropic-on-GCP Vertex, Anthropic direct, et…
Amp's `smart` mode uses Opus 4.7 from multiple redundant providers (Anthropic-on-GCP Vertex, Anthropic direct, etc.), so it's still working despite Anthropic downtime right now.
57.
Can confirm. A lot of our model performance improvements came from bug fixing and data cleaning. Before you have …
Can confirm. A lot of our model performance improvements came from bug fixing and data cleaning. Before you have a trustworthy training infra, you should not fully trust your ablations on model arch or hyperparameters.
58.
1/ Our new #ICML paper targets two practical limitations of uncertainty-based early stopping for reasoning models:
1/ Our new #ICML paper targets two practical limitations of uncertainty-based early stopping for reasoning models: - Setting threshold values for uncertainty signals is hard. - "Stop when confident" principle fails to halt unsolvable proble
59.
i'm working on some sdk design for v2
i'm working on some sdk design for v2 what i keep coming back to is not fully understanding one segment of the usecase every infra company is prepping for "millions of agents to run on us" are you working on a product that needs this? te
60.
I applied my new /rust-unsafe-code-exorcist agent skill to the Bun Rust code:
I applied my new /rust-unsafe-code-exorcist agent skill to the Bun Rust code:
61.
Excited to share SANA-WM: a 2.6B open-source world model for minute-scale 720p video generation. (t.co)
Excited to share SANA-WM: a 2.6B open-source world model for minute-scale 720p video generation. Given one image + text + a 6-DoF camera trajectory, it synthesizes action-controllable 60s worlds on a single GPU. Project: https:// nvlabs.
62.
ported the /goal command from codex to a standalone mcp and slash command for arbitrary agents and harness (for n…
ported the /goal command from codex to a standalone mcp and slash command for arbitrary agents and harness (for now support for claude code and opencode)
63.
@ osdk/react is GA.
@ osdk/react is GA. Hooks for every Ontology primitive, normalized caching, optimistic updates with auto-rollback, and action-driven invalidation so your lists stay in sync without you babysitting them. Check out our GitHub: https:// pa
64.
Just wired our fifth pre-seed investment of 2026:
Just wired our fifth pre-seed investment of 2026: – $750K lead check – Three founders out of Michigan Medicine + Johns Hopkins – AI operating layer for hospital transformation – Live pilots & conversions in some of the largest health syst
65.
IMHO most ppl see CL exclusively as a context length problem and do not weigh the necessity of updating and reaso…
IMHO most ppl see CL exclusively as a context length problem and do not weigh the necessity of updating and reasoning through often conflicting priors to actually utilise what is at hand. Amazing work!
66.
Next wed at the Fonzi office in Williamsburg I'm hosting talks from (x.com)
Next wed at the Fonzi office in Williamsburg I'm hosting talks from @harmonic_ai , @Instacart , @hebbia , and noemica Topics: - yesh (harmonic): feedback loops for agents in hard-to-eval domains - ahsaas (instacart): eval pipelines that
67.
goated article as always! HL came a long way from being just a crypto DEX
goated article as always! HL came a long way from being just a crypto DEX if anyone wants to verify Shaunda's findings themselves, we made $CBRS (and all other Hyperliquid historical data tbh) free, forever
68.
We’re on the hunt to find the best small AI model.
We’re on the hunt to find the best small AI model. Here's the latest AutomationBench scorecard (Based on 2-step workflows with explicit instructions) Most automation in production today doesn't run on the biggest model available. It run
69.
if the idle compute ~14k H200 hours that's 4.98*10^22 FLOPs
if the idle compute ~14k H200 hours that's 4.98*10^22 FLOPs This is ~1% of the FLOPs used for Llama 2 70B (8.3*10^23 FLOPs) lot of idle compute
70.
i had a web server running that allowed me to use codex easily from phone browser.
i had a web server running that allowed me to use codex easily from phone browser. glad to scrap it off today.
71.
Survived my first CNBC hit at 3am HK time. Lineup: (x.com)
Survived my first CNBC hit at 3am HK time. Lineup: @cerebras CFO @BobKomin on IPO day, @OctahedronCap , @SambaNovaAI CEO… and me, talking the first compute futures market with @CMEGroup Thanks @dee_bosa for having me. https://
72.
The OODA loop is a useful model of how to make decisions and execute them. Right now, depending on harness and pa…
The OODA loop is a useful model of how to make decisions and execute them. Right now, depending on harness and part of the stack, the model’s horizon on this OODA loop is a single turn or a single trajectory. There are two ways to think ab
73.
Huge SlateDB release!
Huge SlateDB release! Personal favorite - Splitting an existing database by key range!
74.
> uv add beads
> uv add beads > codex, make plans and write to beads > don’t stop until no other beads are left
75.
New blackboard lecture w (x.com)
New blackboard lecture w @ericjang11 He walks through how to build AlphaGo from scratch, but with modern AI tools. Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the prim
76.
A fun experiment comparing a random step with one gradient step:
A fun experiment comparing a random step with one gradient step: With a small CNN on CIFAR-10, a random step is basically a disaster. (A gradient step is a ~185σ event.) That makes sense if you expect a random direction in R^d to be ~sqrt
77.
Can’t stop thinking about the forthcoming “dreaming” feature from A\.
Can’t stop thinking about the forthcoming “dreaming” feature from A\. Lots to be mined from traces. The Claude Code /insights skill gives a good preview. What it produces is actionable. So, if you can, in an increasingly automated way, de
78.
whos building claude code for checking notifications in slack, gmail, linear, github
whos building claude code for checking notifications in slack, gmail, linear, github
79.
Self-Distilled Agentic Reinforcement Learning (SDAR)
Self-Distilled Agentic Reinforcement Learning (SDAR) SDAR stabilizes multi-turn LLM agent training by gating self-distillation signals within GRPO, yielding +9.4% gains on ALFWorld and significant improvements on WebShop and Search-QA acro
80.
1/ Introducing (t.co)
1/ Introducing http:// purrtrace.com We're building the only explorer purpose-built for HyperEVM. Trace CoreWriter calls, view unified EVM+Core state, and bridging flows- all in one explorer.
81.
fyi if you've tried to use verifiers with tinker in the last couple months and couldn't get it working due to str…
fyi if you've tried to use verifiers with tinker in the last couple months and couldn't get it working due to strange errors - just opened a PR that fixes the issue
82.
Sharing some our learnings from seeing how Claude Code works in LARGE codebases :) (x.com)
Sharing some our learnings from seeing how Claude Code works in LARGE codebases :) Read more on the @claudeai blog: https:// claude.com/blog/how-claud e-code-works-in-large-codebases-best-practices-and-where-to-start …!
83.
No one's talking about how sandbox forking is going to change how multi-agent handoffs work.
No one's talking about how sandbox forking is going to change how multi-agent handoffs work. Right now, when one agent hands work to another, you destroy the VM and start fresh. The context, file state, environment, etc are gone. With for
84.
Stripe Projects turns a network of dev tools into infrastructure that agents can use immediately.
Stripe Projects turns a network of dev tools into infrastructure that agents can use immediately. Provision databases, hosting, auth, analytics, AI, email, observability, and more from the terminal. We're shipping daily - this week: agent
85.
Day-0 vLLM support for Intern-S2-Preview! (x.com)
Day-0 vLLM support for Intern-S2-Preview! Congrats to the @intern_lm team — an open-source scientific multimodal foundation model, with a first take on material crystal structure generation alongside general capabilities. http:// rec
86.
QoL improvement rolling out to Perplexity Computer: When your personal financial data is used in tasks, you can h…
QoL improvement rolling out to Perplexity Computer: When your personal financial data is used in tasks, you can hide it from view in the traceability side panel
87.
This is wild -- how bad do things need to be for a team to use a permissioned function to take a uncollateralized…
This is wild -- how bad do things need to be for a team to use a permissioned function to take a uncollateralized loan against their own user deposits and present it as the DAO fully covered user losses?
88.
here's a quick rundown of how the X algo works for fried attention spans:
here's a quick rundown of how the X algo works for fried attention spans: 1: the algo builds a profile of every viewer - their last 127 engagements, who they follow, what they've muted, so your post has to match the taste pattern of the pe
89.
SWE in heavy agent use teams is basically already bottlenecked on 2 right now
SWE in heavy agent use teams is basically already bottlenecked on 2 right now
90.
Today's autoregressive models generate one token at a time.
Today's autoregressive models generate one token at a time. Mercury 2 generates tokens in parallel. Over 1,000 tok/sec on standard GPUs, at comparable quality to speed-optimized models. Since launch, the community has been showing what d