Backlist — 04 Jun 2026 UTC

VoidZero is joining Cloudflare (x.com)

Cloudflare is bringing the Vite/Vitest/Rollup/Oxc ecosystem closer to its edge platform while keeping the core tools open source

A three-part Pixel 9 zero-click exploit chain (t.co)

Project Zero published a complete chain from media codec RCE to kernel privilege escalation on a current Android flagship

A possible encrypted number station in GPS broadcasts

A public 176-bit GPS navigation message field has carried high-entropy payloads for years, suggesting a world-reachable one-way control or key-distribution channel

Miasma npm supply-chain campaign targets vapi-ai and ai-sdk-ollama

Attackers are using a node-gyp autorun path to compromise popular npm packages without relying on obvious postinstall scripts

Detecting Chrome Incognito via Cache API usage details (t.co)

Writing hundreds of tiny responses to Cache API exposes a measurable storage accounting difference between normal and Incognito Chrome

NVIDIA Nemotron 3 Ultra technical report (t.co)

NVIDIA released a 550B-total, 55B-active hybrid Mamba-attention MoE model with an open post-training stack aimed at agentic workloads

DeepSeek V4 Pro streamed from SSD on a 128GB MacBook

A 1.6T-parameter model running on consumer Apple hardware via SSD streaming changes the practical boundary of local model experimentation

q0: Scaling multi-epoch pretraining when data runs out

Q trains a diverse population of models and aggregates predictions to keep improving across hundreds of epochs instead of saturating a single model

Targeted on-policy self-distillation, explained from an iPhone recording (x.com)

Hint tokens inserted at the exact failure point let a model learn from a bad rollout without regenerating the whole trajectory

Rendering 125GB of Protomaps vector tiles in Three.js

Compute passes and workers let a browser dynamically render a massive vector tile dataset that would normally be treated as server-side GIS infrastructure

No one should be able to order a bioweapon through the mail (x.com)

Mandatory DNA synthesis screening and recordkeeping is a low-cost safety layer against AI-assisted biological weapon design

Jane Street plans its own data center as compute grows scarce

A trading firm building and financing dedicated compute infrastructure shows how AI scarcity is spreading beyond frontier labs and hyperscalers

Deflua: a pure-Elixir Lua 5.3 VM for the BEAM (t.co)

A sandboxed Lua VM inside Elixir enables untrusted user scripts, plugins, formulas, and agent tools without native extensions

Edge.js: full Node.js workloads in a WebAssembly sandbox at the edge

Wasmer built a Docker-free way to run Node workloads in WebAssembly at the edge, pointing toward lighter deployment isolation for serverless apps

1X launches a World Model Lab for embodied AI

1X is betting that general-purpose humanoids need scaled world models trained from physical interaction, not just fine-tuned task policies

Editor’s note: imported_from_x_likes

NewLimit raises $435M to treat cellular aging as disease (x.com)

A $3.1B valuation for NewLimit signals that epigenetic reprogramming and aging biology have moved from speculative longevity research into large-scale company building

Palantir to run the national firearms database for England and Wales (x.com)

A multimillion-pound policing contract gives Palantir a central role in managing firearms data across every police force in England and Wales

The blood cancer that shows China is winning drug discovery (t.co)

Multiple myeloma has become a concrete case study in China’s accelerating ability to turn biotech research into clinically important therapies

Binance lists 7,000 US stocks for non-US users

The interesting part is not brokered stock access but the emergence of self-issued stock tokens that may become portable financial primitives

Analyzing Opus 4.6 tokenflation

A Stanford analysis argues that Opus 4.6 began using far more tokens after launch without a measurable explanation in task demands

MSTR as Luna, STRC as UST on Anchor (x.com)

The analogy frames Strategy’s new instruments as a reflexive Bitcoin-backed structure whose stability depends on narrative strength and market liquidity

National Instruments kernel driver allows arbitrary physical memory access

An EV-signed NI kernel driver used across defense contractors, fabs, NASA test stands, and labs reportedly allows unauthenticated physical memory read and write

Meta’s AI race enters the temporary-tent data center phase

Meta is putting billions of dollars of chips into massive temporary structures powered by off-grid turbines, showing how urgent AI capacity buildouts have become

Relic: a coding agent for 1990s computers

A tiny coding agent that fits on a floppy and runs in 4MB of RAM brings modern agent ideas to machines built before HTTPS was common

Lambda as S3 for compute

AWS Lambda reportedly emerged from the S3 team’s question of whether compute could have a PUT/GET/LIST-like primitive, which explains the shape of serverless invocation

Canada’s AI industry map: 1.2M jobs and 182,000 firms (x.com)

Canada published open data on the full AI supply chain, giving builders and policymakers a concrete map of where the country’s AI economy already exists

Stateful visual language models for comparative reasoning

Adding cross-attention between visual encoder layers targets a common VLM weakness: detecting differences across images, which matters in scientific and medical workflows

Let’s Encrypt’s post-quantum certificate size breakthrough (x.com)

Post-quantum signatures are large enough to hurt global page loads, so practical compression and deployment engineering are becoming central to upgrading web PKI

Multigres: Supabase’s scalable operating system for Postgres

Supabase introduced an alpha Postgres operations layer for high availability today and Vitess-grade horizontal scaling in a future release

We've seen posts circulating about V14 Lite being available or released to some HW3 vehicles.

We've seen posts circulating about V14 Lite being available or released to some HW3 vehicles. As far as we have determined, these posts are entirely fabricated, and we can confirm that no such update has been released to customer vehicles.

employees love to complain about their company, find them. some orgs are large enough to even have unofficial com…

employees love to complain about their company, find them. some orgs are large enough to even have unofficial communities. great opportunities for phishing. 'solutions' for office-specific issues make for A+ pretexts. stick to corp email/IM

*AIRBNB'S CHESKY PLANS NEW AI LAB, IN EARLY STAGES OF FUNDING

*AIRBNB'S CHESKY PLANS NEW AI LAB, IN EARLY STAGES OF FUNDING $ABNB CEO starting a new AI lab to develop AI models. Chesky will remain the ABNB CEO – didn’t have that on the bingo card

SCOOP: Anthropic is gearing up for the public launch of a new version of Mythos, better than Mythos Preview.

SCOOP: Anthropic is gearing up for the public launch of a new version of Mythos, better than Mythos Preview. A checkpoint of the model, codename Oceanus, was made available to red teamers yesterday. These programs typically begin 7 days b

$GOOGL

$GOOGL Scoop: Google's own employees says its AI 'sucks' Internally Google employees are sharing memes about how AI is bad at exact tasks and makes their job harder The people who write the code say the AI they’re using is overhyped

Joined Ramp when we were cramped in a wework. Since then: 4 new offices, 300 > 1500+ people, 140+ hires of my own…

Joined Ramp when we were cramped in a wework. Since then: 4 new offices, 300 > 1500+ people, 140+ hires of my own, and more founding roles than I can count. Today we raised $750M at $44B. Proud doesn't even scratch the surface, but we have

Tijjani Reijnders confirms Dumfries will join Real Madrid: “We have already congratulated him!”.

Tijjani Reijnders confirms Dumfries will join Real Madrid: “We have already congratulated him!”.

- xai is indeed struggling heavily and continues to bleed remaining talent. why would anyone join xai when cursor…

- xai is indeed struggling heavily and continues to bleed remaining talent. why would anyone join xai when cursor people are going to clean house? - cursor isn't that appealing, but a combination of some incredibly high offers and "generou

Can LLMs hack vulnerable apps? I spent the last week trying to find out!

Can LLMs hack vulnerable apps? I spent the last week trying to find out! I made a fake book review app and gave 15 models the APK with the goal: finding a person's private reviews. GPT 5.5 had the best success rate, DeepSeek V4 Pro solve

This is what I’ve spent the past month+ of my life building - designing and implementing the new Railway edge net…

This is what I’ve spent the past month+ of my life building - designing and implementing the new Railway edge network & CDN from the ground up, now serving all of our traffic at 1 million RPS. I wrote about it here!

So did (x.com)

So did @thelinqapp just completely lose their deal? That must have been >15% of their revenue

The response to this has been crazy. So many teams want to move their team docs to (x.com)

The response to this has been crazy. So many teams want to move their team docs to @linear where they work. In the first 24h we added 200+ companies to beta so we now decided to open this for everyone. Start creating team docs, no addit

What does it cost to evaluate 100% of our agent runs?

What does it cost to evaluate 100% of our agent runs? If you're running an LLM-as-judge, the number that comes back is high enough that you end up sampling 10% and moving on. But sampling 10% doesn't really make evaluation cheaper. The

INDIAN REGULATORS HAVE FOUND A PUBLIC COMPANY THAT “FAKED” REVENUE NUMBERS BY $158 BILLION USD

INDIAN REGULATORS HAVE FOUND A PUBLIC COMPANY THAT “FAKED” REVENUE NUMBERS BY $158 BILLION USD THE COMPANY SHOWED REVENUE OF OVER $160 BILLION OVER LAST 5 YEARS BUT 99% OF IT WAS MISLEADING ACCORDING TO THE REGULATORS THE COMPANY IS CALLE

- i can confirm the ant hiring freeze for e5 and below

- i can confirm the ant hiring freeze for e5 and below - oai morale is somewhat low — the valuation flip with ant really seems to have shook some

Currently running a GPU / CPU memory inference of DeepSeek V4 Flash on 1x NVIDIA H200 & 197GB RAM, through KTrans…

Currently running a GPU / CPU memory inference of DeepSeek V4 Flash on 1x NVIDIA H200 & 197GB RAM, through KTransformers + SGLang Opened a PR to integrate DeepSeek V4 tool calling, confirmed working through OpenCode Prompt: build Super Ma

Jamieson seems to have lost the zip

Jamieson seems to have lost the zip

In other news, (x.com)

In other news, @Flipkart , @CRED_club and @gitlab have laid off people with @rubrikInc and tons of @Oracle PPO being revoked.

- tbd has slowed down hiring a little, but hiring at tbd+ continues! what's the difference between tbd and tbd+?

- tbd has slowed down hiring a little, but hiring at tbd+ continues! what's the difference between tbd and tbd+? - thinking machines has become less attractive than early last year, but it's largely stabilized since cofounder departures.

Shared my first trace from (x.com)

Shared my first trace from @NanoClaw_AI to @huggingface yesterday. Very cool! By default, all agents should store their traces on HF (in private) so that you can keep a history of them, analyze them,... & share them and post-train bet

At the same time that Arthur Hayes started selling, another entity related to Andrew Kang ( (x.com)

At the same time that Arthur Hayes started selling, another entity related to Andrew Kang ( @Rewkang ) sold 120k HYPE ($8M) in less than 30 minutes, pushing the $HYPE price down more than 5%. That entity already finished selling his sta

DistIL starts from a simple weakness in RLVR: most of the signal is still one bit at the end.

DistIL starts from a simple weakness in RLVR: most of the signal is still one bit at the end. Instead, it uses richer feedback: execution traces, tool outputs, expert corrections, ground-truth solutions, or model critiques. It replaces re

This may not be broadly known, but if instead of causal attention

This may not be broadly known, but if instead of causal attention yᵢ = xᵢ + attn(norm(x)) you do causal EMA yᵢ = xᵢ + α ∑ⱼ βⁱ⁻ʲxⱼ where α, β are fixed scalars, eg α=0.1, β=0.9, it still works — with a healthy loss curve that converg

EviRank: Evidence-Based Confidence Estimation for LLM-Based Ranking (t.co)

EviRank: Evidence-Based Confidence Estimation for LLM-Based Ranking Estimates position-level confidence for LLM-based ranking by aggregating semantic, attention, and output evidence, with position-aware calibration. https:// arxiv.org/a

Gemma4 12B with Unsloth's Quant on DGX Spark

Gemma4 12B with Unsloth's Quant on DGX Spark Quants: - UD_Q4_K_XL - UD_Q5_K_XL - UD_Q6_K_XL - UD_Q8_K_XL Summary: - Q4: 25.21 tok/s, TTFT 168ms - Q5: 21.7 tok/s, TTFT 182ms - Q6: 17.68 tok/s, TTFT 193.95ms - Q8: 15.22 tok/s, TTFT 221ms

“Trust Region On-Policy Distillation”

“Trust Region On-Policy Distillation” On-policy distillation is powerful, but one bad mismatch between student and teacher can negatively impact the gradients. So this paper's TrOPD only learns where the teacher is reliable, treats outlie

I can’t think of better people other than (x.com)

I can’t think of better people other than @RaghuRaghuram @AnneNeuberger @jkhamehl @astrange and @GEVS94 to lead this global effort. This firm has never seized to pursue the bigger ambition, and building leverage for our founders.

How can we get LLM agents with different capabilities to autonomously self-orchestrate?

How can we get LLM agents with different capabilities to autonomously self-orchestrate? Excited to share Economy of Minds, where agents autonomously learn to cooperate with each other through economic transactions, where agents reward eac

This is so fucking funny

This is so fucking funny Some Chinese hackers have either infiltrated Ant's systems or are part of the red team and are selling Mythos access

BREAKING: We just caught some interesting new stock trades.

BREAKING: We just caught some interesting new stock trades. Representative Josh Gottheimer just filed purchases of: - SanDisk, $SNDK - Micron, $MU - AMD, $AMD - Palo Alto Networks, $PANW Gottheimer sits on the House Subcommitte

Uber employees on-boarding themselves without HR

Uber employees on-boarding themselves without HR

Wow it’s now confirmed (x.com)

Wow it’s now confirmed @tryramp raised a $750M Series F led by @ICONIQCapital at a $44B valuation “Ramp grew TPV ~170% year-over-year in March 2026, the company's highest growth rate in three years” https:// bloomberg.com/news/artic

In v0.21.0, the KV Offload + Hybrid Memory Allocator (HMA) feature was added. Even for models with hybrid attenti…

In v0.21.0, the KV Offload + Hybrid Memory Allocator (HMA) feature was added. Even for models with hybrid attention, you can now offload the KV cache to regular memory, so this is definitely something you should enable. --kv-offloading-size

Building an open-source post-training stack for large language models from first principles.

Building an open-source post-training stack for large language models from first principles. The goal is to understand and implement the systems behind modern reasoning models end-to-end: • SFT • Preference Optimization • RLHF / RLVR • Rew

Miasma, the supply chain campaign that previously compromised 32 (x.com)

Miasma, the supply chain campaign that previously compromised 32 @RedHat packages, is spreading again with a new wave targeting the npm ecosystem. Targets include: - vapi-ai/server-sdk (71k weekly downloads) - ai-sdk-ollama (31k weekly

GPT-5.5 Pro is amazing at almost everything I want it to be, except discussing/reasoning about ideas for retrieva…

GPT-5.5 Pro is amazing at almost everything I want it to be, except discussing/reasoning about ideas for retrieval. It consistently devolves into proposing "multi-facet" representations that make no sense whatsoever. Very weird failure mode

Is there any workaround for getting a better cache hit rate on Gemini 3.1 Pro on Vertex?

Is there any workaround for getting a better cache hit rate on Gemini 3.1 Pro on Vertex? Vertex only seems to have a global endpoint and they keep routing requests to different regions, which reduces our cache hits by almost 50% compared to

Looks like npm packages by (x.com)

Looks like npm packages by @JagReehal got compromised tonight by the same credential-stealing worm that targeted Red Hat npm packages. For example: autotel-devtools@6.1.2 autotel-mcp@29.0.1 Full list of packages: https:// gist.github.c

Building coding agents is mostly harness work. This repo shows the pieces.

Building coding agents is mostly harness work. This repo shows the pieces. Dive into Claude Code is a source-level architectural analysis of Claude Code for builders designing AI agent systems. It helps you move beyond “just call the mode

1/ Two great drops this week, both turning real repos into RL environments: (x.com)

1/ Two great drops this week, both turning real repos into RL environments: - MAI-Thinking-1 ( @MicrosoftAI ) — an in-house SWE env pipeline feeding a frontier RL climb - Repo2RLEnv ( @adithya_s_k ) — open-source, repo → verifiable RL data

Using a generative flow model to solve a difficult signal-processing optimization problem and output deployable F… (t.co)

Using a generative flow model to solve a difficult signal-processing optimization problem and output deployable FIR filters. Nice. Paper: https:// arxiv.org/abs/2606.04570

we rolled out the rust port of bun to claude code internally last night (not on the public builds yet)

we rolled out the rust port of bun to claude code internally last night (not on the public builds yet) I don’t want to jinx it but nobody reported any issues yet and it’s been a day

2014 I pitched at (x.com)

2014 I pitched at @khoslaventures . The partner that was supposed to see us flaked without notice. We got reassigned to someone with no context. He arrived, openly irritated, and sneers “well, I guess you’ve got me” then disappears into

Sophisticated supply chain attack targets CI/CD environments via npm packages using binding.gyp files to bypass s…

Sophisticated supply chain attack targets CI/CD environments via npm packages using binding.gyp files to bypass security audits. Over 286 malicious versions across 56 packages deployed multi-layered encrypted payloads specifically designed

In our prior work ( (t.co)

In our prior work ( http:// arxiv.org/pdf/2509.26030) we showed that Muon outperforms Adam on heavy-tailed knowledge tasks. In this work, we examine Muon's superiority from the perspective of loss curvature. The main takehome message is

And another open-weight release. Nemotron 3 Ultra has an ultra impressive capability:efficiency ratio!

And another open-weight release. Nemotron 3 Ultra has an ultra impressive capability:efficiency ratio! Design-wise, it carries forward the Mamba-2-attention hybrid stack and LatentMoE introduced in the previous Super variant. But everythi

2/ Paper: (t.co)

2/ Paper: https:// arxiv.org/abs/2606.03938 q0 is built on one intuition, motivated by Solomonoff induction: instead of training one perfect model, train a population of diverse models and aggregate predictions. Everything in the algorith

"SWE-bench/ProgramBench are based on publicly-available data, so they're invalid cause the models were trained on…

"SWE-bench/ProgramBench are based on publicly-available data, so they're invalid cause the models were trained on the answers" Nope: 1. Scores are ~0% at first, showing models don't memorize answers. 2. Cheating by post-training on answers

When you run an AI agent today, more than half of what you pay for is the model re-reading the context. (t.co)

When you run an AI agent today, more than half of what you pay for is the model re-reading the context. Analysis: https:// exponentialview.co/p/data-to-star t-your-week-one-ai-task-many-bills …

SSD Streamed Dwarf Start by (x.com)

SSD Streamed Dwarf Start by @anemll , cool demo! Official implementation of streaming is arriving too. DeepSeek Flash should run at ~14 t/s on MacBook m5 max 64GB, DeepSeek PRO should run at 4 t/s on MacBook m5 max 128GB. Those are genera

this is actually a pretty cool demo that seems to have gone under appreciated

this is actually a pretty cool demo that seems to have gone under appreciated you should try this on your own api, if agents can use your api / mcp / cli to recreate your entire product it isn't agent accessible

MetaPoint is a clean fix for spatial control in image generation: make the coordinate itself a token.

MetaPoint is a clean fix for spatial control in image generation: make the coordinate itself a token. It uses the model’s existing positional encoding instead of new architecture, large coordinate vocabularies, or custom attention masks.

misc thoughts from writing some code by hand for the first time in a bit:

misc thoughts from writing some code by hand for the first time in a bit: - there are so many microdecisions that you make while manually coding that get lost when looking at a plan - 0 skill atrophy, immediately got back into being able t

We're fixing a codex bug today that was causing us to undercount tokens being served to some Pro and Plus account…

We're fixing a codex bug today that was causing us to undercount tokens being served to some Pro and Plus accounts by a small amount. This impacted < 15% of accounts. Not the kind of bug you want us to fix, but didn't want to do this silen

Some cool work that I co-mentored with (x.com)

Some cool work that I co-mentored with @NeelNanda5 I recommend the appendix section on practical AO evaluation details. In particular, consensus sampling significantly reduces hallucinations, and eval performance majorly improves with

I analyzed Trend Micro Deep Security Agent for Linux and found that a local event storm can force bmhook/tmhook r… (t.co)

I analyzed Trend Micro Deep Security Agent for Linux and found that a local event storm can force bmhook/tmhook reload cycles, opening a repeatable temporary protection bypass window. Full write-up: https:// matheuzsecurity.github.io/hac

Pinterest announced this morning they will pay AWS $4 billion for cloud services through 2031. Largest infrastruc…

Pinterest announced this morning they will pay AWS $4 billion for cloud services through 2031. Largest infrastructure commitment in the history of the company.

Second big release from us today: Nemotron-3.5-ASR-Streaming!

Second big release from us today: Nemotron-3.5-ASR-Streaming! 40 languages 80ms - 1s controllable latency 240 - 2400 concurrent streams on 1xH100 FastConformer Cache-Aware RNN-T architecture

This is wild… (x.com)

This is wild… @voidzerodev is joining @Cloudflare !!!! I knew I made a great decision two months ago but it just keeps getting better and better!

We want to work with kernel developers to help them publish their cool kernels on the (x.com)

We want to work with kernel developers to help them publish their cool kernels on the @huggingface Hub via Kernels. This has several advantages: * A consistent build structure * Extreme ease of use * Standardized distribution * Reprodu

Highlighting recent advances in multi-GPU and tensor parallel support in llama.cpp

Highlighting recent advances in multi-GPU and tensor parallel support in llama.cpp Over the last few months llama.cpp maintainers and engineers from NVIDIA collaborated to improve the multi-GPU performance in ggml. This resulted in signif