1.
AI-driven formal proof search for open Erdős problems
A Google DeepMind paper describes autonomous Lean-based proof search resolving open Erdős problems at per-problem costs of only a few hundred dollars
Backlist — 24 May 2026 UTC
Chose one representative per repeated story where possible, especially around AI math results, agent tooling, and semiconductor rumors.
2.
WhatsApp chat databases stored plaintext in a Meta app-group container
On iOS and macOS, WhatsApp chat databases can sit plaintext in a Meta app-group container, making same-developer app boundaries much weaker than users expect
3.
TrapDoor supply-chain attack hits npm, PyPI, and Crates.io
TrapDoor shows active attackers coordinating malicious releases across npm, PyPI, and Crates.io rather than targeting a single package ecosystem
4.
Microsoft proposes a ratepayer-protection tariff for large data-center loads (x.com)
Microsoft’s proposed Nevada tariff would assign new data-center grid costs to the large customer that causes them, a concrete model for AI power load politics
5.
DeltaBox: faster checkpointing for RL tree search and agent fan-out
DeltaBox attacks a systems bottleneck for agent tree search: cheap process and filesystem checkpointing without copying an entire sandbox every branch
6.
Bun now fuzzes every language parser 24/7 and routes repros to Claude
Bun is using continuous parser fuzzing plus minimized repros routed to Claude, turning crash discovery into a mostly automated repair pipeline
7.
FedWire access for fintechs: useful, but not a panacea (t.co)
Expanded FedWire access helps fintechs only if they can solve the hard part: funding and managing settlement accounts, not just connecting to the rail
8.
Stablecoins EURR and USDR exploited through a 1-of-3 minting multisig (x.com)
A $10M exploit hit EURR and USDR through a 1-of-3 minting multisig, showing how stablecoin designs still collapse around key-management assumptions
9.
vLLM bans contributor after fake “PR training” issue (t.co)
vLLM’s fake-issue PR episode shows open-source maintainers now have to defend review time against resume-driven AI slop
10.
Microsoft open-sources 86-DOS 1.00
The newly opened 86-DOS 1.00 source is the pre-IBM PC DOS codebase behind the compatibility lineage that shaped personal computing
11.
Marin preregistered the loss for a 1e23 FLOPs MoE run
The Marin team preregistered a large MoE training loss before launch and beat it, making scaling-law forecasting auditable instead of retrospective
12.
China’s HTR-PM pebble-bed reactor module heads toward commercial service
China’s HTR-PM pebble-bed module moving toward 2027 commercial service is a live test of whether advanced nuclear can be built repeatably and economically
13.
Vietnam’s manufacturing exports are 90% of GDP, but value add lags
Vietnam’s export boom masks a development trap: manufacturing exports reach 90% of GDP while high-value components are still imported
14.
When SGLang OOMs, what exactly runs out of memory?
SGLang OOMs are not one failure mode; serving memory can disappear into weights, KV cache, activations, fragmentation, queues, or scheduler choices
15.
Using Hopper and MCP to inspect Apple platform frameworks (t.co)
Pairing Hopper with an MCP server gives Apple engineers a practical workflow for interrogating private frameworks and dyld shared caches with code agents
16.
Hyper: an API framework as source, not a dependency (x.com)
Hyper experiments with a shadcn-style API framework that is copied into your repo, giving routes OpenAPI, typed clients, and MCP without a runtime dependency
17.
A DNS allowlist for in-flight WiFi login portals (t.co)
A tiny DNS allowlist for in-flight WiFi portals solves a real travel problem for users with VPNs, DNS filters, and locked-down company devices
18.
A homegrown Toyota Production System at SendCutSend (x.com)
SendCutSend’s homegrown Toyota Production System shows modern manufacturing advantage coming from operational discipline, not just software wrappers
19.
Why Japanese firms do many things
Milgrom-Roberts complementarity explains why Japanese firms historically combine low hierarchy, broad worker tasks, lifetime employment, and unrelated product lines
20.
SwiftUI to AppKit rewrites can still yield 3x speedups
A SwiftUI-to-AppKit rewrite producing 3x speedups is a useful counterexample to simplistic framework debates: abstractions can be fast, but the escape hatch still matters
21.
A conditional protein diffusion model (t.co)
Conditional protein diffusion points generative modeling at protein design, where the artifact is a physical molecule rather than an image or text sample
22.
PhysX-Omni: sim-ready generation for physical AI (t.co)
PhysX-Omni packages sim-ready generation, datasets, and benchmarks for rigid, deformable, and articulated objects, addressing a missing layer in robotics data
23.
TurboQuant+ shrinks KV cache memory 4.75x (x.com)
TurboQuant+ shrinks KV cache memory 4.75x with 3-bit quantization across CUDA and Metal while preserving near-fp8 top-5 behavior
24.
A long-awaited xterm.js rendering fix (t.co)
An xterm.js rendering fix is a reminder that terminal correctness still depends on unglamorous edge cases that affect thousands of developer tools
25.
Code quality as CI and observability, not aesthetics
Framing code quality as CI and observability shifts the debate from aesthetics to whether failures, performance, and regressions are visible
26.
Austin’s HOME initiative enables three homes on single-family lots
Austin’s HOME initiative shows a concrete zoning change: allowing three homes on single-family lots can create family-sized infill without megaprojects
27.
Show HN: an SF fog webcam aggregator (t.co)
A simple SF webcam aggregator turns hyperlocal fog and microclimates into usable trip-planning data that forecasts often miss
28.
Next up, Anthropic on SWE Bench Pro.
Next up, Anthropic on SWE Bench Pro. This is where we see some bigger jumps rather than incremental. Opus 4.6 scored 53.4%, Opus 4.7 hit 64.3%, and Mythos Preview jumped to 77.8%. • Opus 4.6: 53.4% • Opus 4.7: 64.3% • Mythos Preview: 77.8
29.
"Our most capable agent autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundr…
"Our most capable agent autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundred dollars, proved 44/492 OEIS conjectures, and is being deployed in combinatorics, optimization, graph theory, algebraic geome
30.
OMG!
OMG! 97.8 TPS single 3090 qwen3.6 27b dense. (P.S. No overclocking yet) Recipe cooking at
31.
Our inference stack, optimized for Blackwells, with a novel attention kernel and many new optimizations has start… (x.com)
Our inference stack, optimized for Blackwells, with a novel attention kernel and many new optimizations has started rolling out! It's already charting on Artificial Analysis, eg: #1 speed and latency for @Kimi_Moonshot Kimi 2.6. #1 on l
32.
context-mode keeps raw tool output out of your AI agent's context window.
context-mode keeps raw tool output out of your AI agent's context window. 98% reduction · 15 platforms · 15.5K stars · 1.1K forks · No telemetry Used across teams at: Microsoft · Google · Meta · Amazon · IBM · NVIDIA · ByteDance · Stripe
33.
One of the deepest realizations I’m having about systems/GPU/distributed computing:
One of the deepest realizations I’m having about systems/GPU/distributed computing: "Hardware does NOT create parallelism." It only consumes parallelism that already exists in the data/workload. That changes everything. Parallelism fund
34.
When SGLang OOMs, What Exactly Runs Out of Memory?
When SGLang OOMs, What Exactly Runs Out of Memory? Around two months into fully developing SGLang Omni, roughly this April, we got a brand-new H100 on top of the H200 development machine and the H20 CI machine we already had. That meant one
35.
Happy to share our new ICML oral in Pretraining: OPUS!
Happy to share our new ICML oral in Pretraining: OPUS! It tells: which tokens should the model train on at each step? Instead of static data filtering, OPUS dynamically selects tokens based on the optimizer-shaped updates. Less “more d
36.
This. Understand the code you're responsible for.
This. Understand the code you're responsible for. "My AI agents one-shotted this app overnight and I didn't read any code and just shipped it" is not impressive; it's irresponsible.
37.
Александр Рыжков (Kaggle Grandmaster, LightAutoML team lead) published a comprehensive Russian-language DVC tutor… (t.co)
Александр Рыжков (Kaggle Grandmaster, LightAutoML team lead) published a comprehensive Russian-language DVC tutorial covering data versioning, S3 & Google Drive backends, DAG pipelines, and when NOT to use DVC. https:// hubs.la/Q04fRcSy
38.
yeah that's it: each link is it's own DO for:
yeah that's it: each link is it's own DO for: - upload state - capability auth - per-share limits - expiry/cleanup lifecycle - concurrency control R2 holds the file bytes. There is also a separate DO for deployment-wide quota limits
39.
slime Source Code Walkthrough: SGLang-Native Inference Architecture (Part II)
slime Source Code Walkthrough: SGLang-Native Inference Architecture (Part II) Evaluation & insights from Zhihu contributor 鲸饮未吞海 Inference Data Plane Chapter 2 explained how the system is launched. Now we switch perspective: after every
40.
[1/4] The human eye doesn't process every single pixel of a video continuously—it focuses on what changes.
[1/4] The human eye doesn't process every single pixel of a video continuously—it focuses on what changes. So why are our video AI models wasting compute on redundant frames? Introducing Swift Sampling: a test-time technique inspired by t
41.
If you have an Nvidia RTX 4090 --ddtree-budget 36 is the best configuration that buys you 2.5x speed up during de… (t.co)
If you have an Nvidia RTX 4090 --ddtree-budget 36 is the best configuration that buys you 2.5x speed up during decoding for Qwen3.6_27B. Thanks for the benchmark https:// github.com/1TommyCheung
42.
New NanoGPT Speedrun WR at 81.8 (-2.6s) from @.Lisennlp on Github with MUDD skip connections, an expressive and e…
New NanoGPT Speedrun WR at 81.8 (-2.6s) from @.Lisennlp on Github with MUDD skip connections, an expressive and efficient mechanism for data dependent skips! Instead of a learned scalar or sigmoid(linear) gate, MUDD uses a 64 neuron 'MLP' t
43.
Qwen inference team is super great — they achieved 540 TPS on TokenSpeed for agentic workloads (t.co)
Qwen inference team is super great — they achieved 540 TPS on TokenSpeed for agentic workloads Looking forward to them sharing more optimization details soon. Stay tuned. https:// github.com/lightseekorg/t okenspeed … Enjoy!
44.
New: Traces for Mixedbread agentic search
New: Traces for Mixedbread agentic search See every search call an agent makes directly in the dashboard, and tune instructions for better retrieval quality.
45.
New in emulate v0.6: way more Slack
New in emulate v0.6: way more Slack Emulate Slack in CI + agent sandboxes CLI or Next.js adapter Messages Threads Channels DMs MPIMs OAuth Inspector SDK tests Scopes Profiles Presence Files Uploads Pins Bookmarks Apps Modals Webhooks Eve
46.
Workbench - open-source BullMQ dashboard, drop-in for any Node backend.
Workbench - open-source BullMQ dashboard, drop-in for any Node backend. Flows, metrics, schedulers, search. MIT. Link
47.
Think ⨉ Skills (t.co)
Think ⨉ Skills --- https:// github.com/cloudflare/age nts/pull/1584 … - supports the https:// agentskills.io spec - load via local filesystem/codebase, or r2 (git coming soon? maybe) - configurable permission model - working js/python/ba
48.
@saturdayrobotic (x.com)
@saturdayrobotic & World Model Reading Club 09, Part 1 Recap, @CVPR Warm-up: keynote @tommiekerssies , hosts @junfanzhu98 , @aurorafeng_01 , @zoeytzh A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tok
49.
What's the secret sauce behind leaning attractors?
What's the secret sauce behind leaning attractors? Stable & amortized optimization. Here is the controlled study of how we built EqR!
50.
How did Tensor Cores massively increase the throughput of Nvidia chips? (x.com)
How did Tensor Cores massively increase the throughput of Nvidia chips? @reinerpope explains the fundamental idea, systolic arrays:
51.
There is now a smarter way to pick data for training LLMs!
There is now a smarter way to pick data for training LLMs! Enter OPUS! This is an ICML Oral paper from SJTU, Alibaba, UW–Madison, UIUC, and Mila - Quebec AI Institute. The proposed method dynamically and intelligently selects the most im
52.
Fast browser agents are getting cheap.
Fast browser agents are getting cheap. Composer 2.5 in Pi drove Sauce Demo checkout in a headed browser and placed the demo order. 39s end-to-end 28 browser tool calls ~7.7k Pi-estimated model tokens pi-agent-browser-native v0.2.33 What
53.
>You pay per token, in & out. Input tokens are cheaper if they’re cached, but it costs to write them to a cache, …
>You pay per token, in & out. Input tokens are cheaper if they’re cached, but it costs to write them to a cache, and the cache expires when we decide. If you add an image, it may invalidate the cache. We decide how many tokens to output, an
54.
the mental model that finally made agents click: a stateless reducer on a stream
the mental model that finally made agents click: a stateless reducer on a stream
55.
I’ve left Google DeepMind.
I’ve left Google DeepMind. The last two years have been an incredible whirlwind. A couple years ago, I joined a small startup called Codeium. There, I got to ship Windsurf, train SWE-1 (a frontier agentic coding model), go to DeepMind in
56.
NYC weekends really do feel rainier... I ran more data and built a full data investigation to see if the pattern …
NYC weekends really do feel rainier... I ran more data and built a full data investigation to see if the pattern is real. Across 2,192 days (6yrs now!) of weather data, NYC rain rates by day of week look like this: Mon: 27.8% Tue: 32.9% W
57.
Paper here: (t.co)
Paper here: https:// arxiv.org/pdf/2502.12170. The MUDD coefficients are used for many purposes, such as routing multiple layers into future attention values, modulating the value embedding, modulating the bigram embedding, etc. (Delay on
58.
Send help! The first part of my blog seires on "how to profile like a noob" is going out of bounds now.
Send help! The first part of my blog seires on "how to profile like a noob" is going out of bounds now. I am here explaining why there was a cudaOccupancy runtime call before cudaLaunch for `aten:mm` and not `aten:add` This is fascinating
59.
I got bombarded with fraudulent orders.
I got bombarded with fraudulent orders. Recently someone (competitor) hired some BH service that puts in fake orders using a bunch of different credit cards. Then they chargeback and get your SP banned. Best way to avoid this is to setup
60.
I have a stacktrace right here. This is stable diffusion in pytorch, right after flash-attention was updated. The…
I have a stacktrace right here. This is stable diffusion in pytorch, right after flash-attention was updated. The only difference between clean wholesome image generation and this compute-sanitizer IMA was that flash-attention upgrade remov
61.
A lot of automated crash reports on Windows have anti-cheat DLLs in the stack trace. Very hard for us to do anyth…
A lot of automated crash reports on Windows have anti-cheat DLLs in the stack trace. Very hard for us to do anything about it. Chrome has a hardcoded list of these where they basically ignore crash reports from.
62.
Pet peeve: "I value correctness over performance".
Pet peeve: "I value correctness over performance". It sounds kinda true (FKs, ACID, and writing to something other than /dev/null can be slower). But when the type of system you are building is known (e.g. Postgres), the tradeoff is bo
63.
“Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units”
“Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units” One of the papers I personally found very promising earlier this year just got accepted as an ICML Oral. huge congrats to the authors The core objec
64.
new aithy website (t.co)
new aithy website https:// github.com/dosco/aithy - better secure sandbox - artifact management - tons of built in skills - new multi-service arch., - packaged deployer - mesh networking - usage tracking - permissions system - memories -
65.
minecraft lobby from first principles
minecraft lobby from first principles
66.
After our Perception Tokens paper, we asked: are models truly reasoning over perception tokens or do they just be… (x.com)
After our Perception Tokens paper, we asked: are models truly reasoning over perception tokens or do they just benefit from extra reasoning budget? And why discrete tokens instead of richer continuous ones? With @JackZhang970191 we answer
67.
Scheduled Tasks for Project Think (t.co)
Scheduled Tasks for Project Think https:// github.com/cloudflare/age nts/pull/1585 … - use cron patterns or a DSL - run a prompt (or regular code, coming soon) - ... that's it. simple.
68.
added a new exposure catalog to bumblebee (t.co)
added a new exposure catalog to bumblebee https:// github.com/perplexityai/b umblebee …
69.
deepseek is the model to watch. amazing value for what it does for batch work. (x.com)
deepseek is the model to watch. amazing value for what it does for batch work. also @FireworksAI_HQ your pricing needs to be updated. 4x more expensive than milking the tokens direct from the creature.
70.
i was honestly surprised when i saw the qwen 3.5 4b working with ax agent in aithy it was writing javascript code…
i was honestly surprised when i saw the qwen 3.5 4b working with ax agent in aithy it was writing javascript code in a repl and bash code in a sandbox to do stuff. and all of this running on my m1 pro.
71.
Be very careful clicking on any links, etc. in the official Cloudflare Discord server right now. Crypto scams and…
Be very careful clicking on any links, etc. in the official Cloudflare Discord server right now. Crypto scams and more remain unmoderated for days sometimes. Cloudflare disbanded their Community Champs program - folks who moderated the ser
72.
For all you terminal nerds, a quick sneak peek at a QoL improvement shipping in the next version of (x.com)
For all you terminal nerds, a quick sneak peek at a QoL improvement shipping in the next version of @evo__hq You can now track your autoresearch runs right inside your terminal, no need to open the dashboard (which evo already ships wit
73.
This paper reframes data efficiency as Data Value Density: more training effect per unit of data, through selecti…
This paper reframes data efficiency as Data Value Density: more training effect per unit of data, through selection, removal, scheduling, mixing, augmentation, distillation, and evolution. From a Deep Manifold lens, DVD is boundary-conditi
74.
I really want to learn more about inference engineering, projects like (x.com)
I really want to learn more about inference engineering, projects like @sgl_project , @vllm_project really excites me trying to push the limits of a gpu. Really want to be good enough to contribute to these OSS projects. Starting today
75.
Check RACO, accepted as an 𝗢𝗿𝗮𝗹 paper to #ICML2026 (𝗧𝗼𝗽 𝟬.𝟳%) (t.co)
Check RACO, accepted as an 𝗢𝗿𝗮𝗹 paper to #ICML2026 (𝗧𝗼𝗽 𝟬.𝟳%) we propose a new conflict-averse optimization scheme for LLM multi-objective finetuning, with counterintuitive theoretical acceleration and better empirical pareto fron
76.
Tired of spending weeks of SFM reconstruction? (x.com)
Tired of spending weeks of SFM reconstruction? Try VGGT-horseshoe Plugin in @lichtfeldstudio to get pose and dense point + sky segmentation in 20 seconds.
77.
assuming 100,000 human hours spent thinking about this problem (napkin)
assuming 100,000 human hours spent thinking about this problem (napkin) worldwide weighted average of $20/hr for maths research, that $2,000,000 the cost of frontier maths research just fell at least 2,000-fold.
78.
does anyone at oai know the time it took human researchers to verify the proof
does anyone at oai know the time it took human researchers to verify the proof
79.
Control strategy for large scale bio datasets serving as substrate for AI is its own science (and one I’d suggest…
Control strategy for large scale bio datasets serving as substrate for AI is its own science (and one I’d suggest we played a very meaningful role in pioneering). Can’t agree more with Ron. If you see controls in a row, column or on the ed
80.
It doesn't have to be a choice between fast-and-fragile or slow-but-secure but if you have an application you car…
It doesn't have to be a choice between fast-and-fragile or slow-but-secure but if you have an application you care about you need infra building blocks that guide your project towards simplicity. Systems design is more important than ever.
81.
update — released blindcache v0.2 with semantic search.
update — released blindcache v0.2 with semantic search. embeddings happen locally via Xenova/all-MiniLM-L6-v2 in-process, so your text never leaves the SDK to be embedded. mem0, letta, zep all send your plaintext to openai's embedding api
82.
Maple gives you a lot of different options to visualize and explore a trace for any use case.
Maple gives you a lot of different options to visualize and explore a trace for any use case. Debugging an error? Use the waterfall. Debugging perfomance issues? Use the timeline. Exploring the executing order or debugging retries? Use th
83.
Ever wished you could adapt your EP size on the fly for fault tolerance of scaling purposes?
Ever wished you could adapt your EP size on the fly for fault tolerance of scaling purposes? Now you can with NIXL-EP! Check it out:
84.
The International Economic Review has just published a wonderful paper by the late Marcus Hagedorn (with special …
The International Economic Review has just published a wonderful paper by the late Marcus Hagedorn (with special thanks to Iourii Manovskii) on A Demand Theory of the Price Level. It is available through @WileyEconomics here:
85.
introducing >molly
introducing >molly a terminal-native discord client discord is slow, heavy, and resource hungry - a browser wrapped in electron, burning your ram just to chat. molly fixes that open your terminal, type molly, and you're in discord. no br
86.
My newest gbrain-evals just dropped - this is how gbrain does vs other options. (t.co)
My newest gbrain-evals just dropped - this is how gbrain does vs other options. http:// ZeroEntropy.dev is SOTA for reranking and embedding cost, speed, and retrieval success. GBrain beats MemPalace by 1% on LongMemEval and beats Vector R
87.
Data synthesis is not only about generating more data, but about generating the right data efficiently. How to sy…
Data synthesis is not only about generating more data, but about generating the right data efficiently. How to synthesize effective and diverse training data in a cost-efficient way is a key question for LLM improvement. This work provides
88.
A way we have started framing this lately in our team is that all tech debt fixes must be aimed to make the agent…
A way we have started framing this lately in our team is that all tech debt fixes must be aimed to make the agents better. Everything else is a waste.
89.
Excited to share that our paper “Closed-Form Concept Erasure via Double Projections” has been accepted to #CVPR2026!
Excited to share that our paper “Closed-Form Concept Erasure via Double Projections” has been accepted to #CVPR2026! As generative models become increasingly powerful, concept erasure is emerging as an important problem for trustworthy and
90.
In the agentic economy, once agents start moving value, I believe the most interesting question is: which infra t…
In the agentic economy, once agents start moving value, I believe the most interesting question is: which infra the agents run on? Rails decide everything here & I think it will condense to a core set of requirments: - *Throughput* is the