No, OpenAI didn't solve P vs NP.
The claim is narrower. Its model found an algorithm that refutes the Exponential Time Hypothesis, a 1999 conjecture that 3-SAT needs exponential time. There's a Lean proof, and two of ETH's authors are reviewing it.
https://t.co/KgYjfmwG8w
last30days is trending on GitHub. It's an agent skill: give it a topic or a person, and it searches Reddit, HN, Polymarket and GitHub with no API keys, ranks posts by upvotes and likes, and writes one brief.
X and YouTube need your own keys.
https://t.co/mAFneejYm1
Gemini 4 Neon is out at $3 input and $15 output per 1M tokens, with a 1M-token context.
The number I'd look at: on SWE-sweep, the same model scores 12.8% in mini-SWE and 39.0% in Google Antigravity.
Compare coding agents with their tooling included.
https://t.co/nVzG3ZVHq6
An nginx bug from 2015 lets an attacker run code on the server with no login. It's in 1.9.5 through 1.31.8 when HTTP/2 is on. Wiz got RCE on a box serving only static files.
Fixed in 1.30.7 and 1.31.9. Can't upgrade today? Set http2 off; for now.
https://t.co/cl9TBVSAbg
AG-UI is trending on GitHub. It's a protocol between an agent and a frontend, at 1.0 with a frozen spec.
The agent sends typed events: text deltas, tool calls, state updates, pauses for humans. Any HTTP endpoint that takes RunAgentInput can be an agent.
https://t.co/MbJqY12rfk
GPT-6.2 is out as two models, and the cheaper one is the coding story.
On Cognition's FrontierCode 2.0, Sol scores within 0.5 points of Astra at about a fifth of the cost.
Astra API costs $8 input, $0.40 cached input and $40 output per 1M tokens.
https://t.co/Rd6emg2Ej5
Java4Shell is a memory-corruption bug in Java's Base64 decoder.
It hits OpenJDK 17.0.6+ on x86-64 CPUs with AVX-512 VBMI. Decoding 64 "A"s plus "=" writes 62 bytes past the array.
Patch now. Stopgap flag: -XX:-UseBASE64Intrinsics
https://t.co/3vurYzpOkM
Run your own mail server? Spamhaus is rolling out a new rule this month: every sending IP needs a valid HELO, a PTR record, and an A record that resolves back to that IP.
Quick check:
dig +short -x YOUR_IP
dig +short THAT_HOSTNAME
Both should agree.
https://t.co/WVj9gVnLz8
A Windows 11 DLL checks for codex.exe, claude.exe, cursor.exe and copilot.exe.
Xusheng Li renamed cmd.exe to codex.exe and ran it. Windows then queued a Codex usage bit to send to Microsoft.
It's optional diagnostic data, on by default in his US setup.
https://t.co/Fhbv7LrFmL
NVIDIA's SkillEvaluator is trending on GitHub. Its harness-check says 2 of 6 agents won't load my test skill.
Claude Code skips .agents/skills. Copilot CLI drops any skill whose description is over 1,024 UTF-16 units. Mine had 1,030.
https://t.co/HRbO8p4Uz0
Turbo Pascal 3.02 for DOS, the compiler and the editor together, was 39,731 bytes.
GNU true on my Debian box is 43,432 bytes. Its whole job is to exit with code 0.
jQuery 4's minified build is 78,748 bytes, about two Turbo Pascals.
https://t.co/TK6HYMMxom
Qwen4-Max is out. The number I'd look at first is cost, not arena rank.
On Alibaba's own Work360 eval, average cost per task fell from $4.43 to $1.67 vs Qwen3.8-Max, with 24-35% fewer tokens. On Terminal-Bench 4.0 it used about 48% fewer.
https://t.co/TW2xI3qlej
Chrome DevTools MCP is trending on GitHub. It gives a coding agent a real Chrome: perf traces, network requests, console logs with source maps.
It also ships a plain CLI. Google usage stats are on by default. --no-usage-statistics turns them off.
https://t.co/RU2fj3Eu9U
Woz wrote the Apple 1's BASIC before he built the computer.
He had no assembler, so he wrote 6502 code on paper and looked up each opcode's hex by hand. He later said the Apple II's BASIC was "one handwritten copy, all handwritten, all hand-assembled."
https://t.co/iO4sMK4tOR
Michigan researchers found LLMs used 1.4x to 180x more energy on non-English text, mostly from extra tokens.
I tried my own language on the same UN declaration. With GPT-4's tokenizer, Georgian takes 10.9x the tokens of English. With GPT-4o's, 1.7x.
https://t.co/k67O2c1Mqx
Alibaba open-sourced the AI code review CLI it used internally for two years.
Plain code picks the files, bundles them, and pins where each comment lands. The agent only judges the code. They report 1/9 of Claude Code's tokens, with lower recall.
https://t.co/hCNsDUoMY2
Box-drawing borders cost a TUI a full row or column, because they're text.
Ratatui's maintainer patched Ghostty to draw borders on cell edges, stored per cell like color.
He used DCS, not APC. Kermit 95 has run APC payloads as commands since the 90s.
https://t.co/RjQlsBdlZT
ggml, the tensor library under llama.cpp, hit 1.0 yesterday, about four years after the repo was created.
The release drops the deprecated API and adds FP8 support, MoE fusion on CUDA and Metal, and a Vulkan FlashAttention prefill path.
https://t.co/sQ0NnkeEhy
Tailscale's tailcat is trending on GitHub. It's netcat over Tailscale's WireGuard data plane, with no account and no root.
One side gets a short address and the other pipes into it. It starts on a DERP relay, then goes direct over UDP when it can.
https://t.co/nWJ3FXLEUU
Rust's never type took 10 years and 5 failed stabilization attempts. Rust 1.100 shipped it yesterday.
You can now write Result<T, !> for a result that can't fail, and let Ok(v) = res compiles with no Err arm. Infallible is now just an alias for !.
https://t.co/u2bo0TSvwp
Google's new fast model beat one of its frontier models. Gemini 4 Hydrogen scores 51 on the Artificial Analysis index vs Argon's 50.
Paired with Carbon in Antigravity, it hit near-Carbon success on Terminal-Bench 4.0 with 24% less quota.
https://t.co/tDl9EeLs6X