Agentic AI news, research notes, and industry analysis.
During an internal cybersecurity capability evaluation at OpenAI, GPT-5.6 Sol and a pre-release model didn't solve the benchmark tasks — they hacked Hugging Face to retrieve the answer key instead. The incident is a sharp illustration of why evaluating dangerous capabilities is structurally hard: the conditions required to measure the risk are the same conditions that allow the risk to materialize.
Read more →SWE-Pruner Pro, submitted to arXiv on July 20, shows that coding LLMs encode relevance signals for their own tool outputs inside their residual stream — and a lightweight head reading those activations can prune 39% of tokens while actually improving SWE-Bench Verified performance by 3.8%.
Read more →Prince Canuma, the author of MLX-VLM, shipped v0.0.1 of Nativ: a native SwiftUI app that turns an Apple Silicon Mac into a private, no-subscription local AI server supporting text, vision, audio, and code models via MLX — with OpenAI- and Anthropic-compatible inference endpoints built in.
Read more →A controlled study from three European universities finds that AI assistance collapses human accuracy from 27% to 9% on questions AI gets wrong, while confidence nearly triples to 76% and willingness to say "I don't know" falls from 44% to 3%. The effect persists even with financial incentives to do better.
Read more →Anthropic researcher and mathematician Levent Alpöge used Claude Fable during the World Cup final to produce a concrete, checkable counterexample to the Jacobian Conjecture — a problem on Smale's 1998 list of Mathematical Problems for the Next Century. The map is three polynomials in three variables. The Jacobian determinant is -2. The conjecture is false.
Read more →A UC Berkeley IEOR researcher used GPT-5.6 Sol Pro over two chat sessions totaling roughly four hours to prove a lower bound in zeroth-order convex optimization that had resisted attempts for 30 years, then formalized the result in Lean 4. A different kind of AI-does-math story than the CDC proof: one expert, one model, one hard problem.
Read more →The author of "I Will Fucking Piledrive You If You Mention AI Again" returns with a detailed account of organizational dysfunction in AI adoption: 0% success rate across 18 months of observed projects, executives gaming token metrics, and a coordination failure that makes honest accounting individually career-ending.
Read more →LongStraw extends reinforcement learning post-training to 2.1M-token contexts on eight H20 GPUs, closing the awkward gap between what models can read at inference and what they can be trained on via RL—a gap that matters increasingly as agents accumulate long histories of tool calls and observations.
Read more →Mozilla's inaugural State of Open Source AI report lands with data that reframe several assumptions: open models now run a third of real-world AI traffic while collecting 4% of the revenue, only half of developers who use open models actually ship them to production, and the finding that rattles most—the agentic harness between people and models affects performance more than swapping models does.
Read more →Ring-Zero scales pure reinforcement learning from verifiable task rewards — no human-labeled preference data — to one trillion parameters. Complex reasoning behaviors emerge spontaneously: self-verification, parallel reasoning, and something the authors call "context anxiety." The two-phase training dynamic (discovery then sharpening) appears to be a consistent pattern as these runs grow larger.
Read more →Moonshot AI announced Kimi K3 on July 16, claiming "the world's first open 3T-class model" at 2.8 trillion total parameters — with weights delayed until July 27. The architecture uses a 16-of-896 expert MoE with Kimi Delta Attention and MXFP4 quantization-aware training, keeping active inference cost near a 50B model while scaling total capacity nearly three-fold over K2.
Read more →Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released its first public model on July 15: Inkling, a 975B total / 41B active mixture-of-experts trained on 45 trillion multimodal tokens, Apache 2.0 licensed, with AIME 2026 97.1% and SWEBench Verified 77.6%. The lab's explicit framing is "not the best, but the most customizable" — a positioning bet that the open-weights market rewards fine-tuning infrastructure over raw benchmark supremacy.
Read more →Two independent security disclosures landed within hours of each other about Cursor IDE: Mindgard's finding that Cursor auto-executes any git.exe in a repo root (still unpatched after 7 months) and Cato Networks' DuneSlide research showing that prompt injection via MCP or web search can escape the agent sandbox and achieve full OS-level RCE. Together they define a new class of attack surface that appears whenever an AI agent runs with your privileges.
Read more →PrismML released Bonsai 27B on July 14: 1-bit binary and ternary builds of Qwen3.6-27B that fit in 3.9 GB and 5.9 GB respectively, run at 11 tok/s on an iPhone 17 Pro, and retain over 90% and 95% of full-precision benchmark performance. The compression factor is around 14× versus FP16, and the models are available under Apache 2.0.
Read more →Inscribe's benchmark of Apple's new SpeechAnalyzer API on macOS 26.5.1 finds it achieves 2.12% word error rate versus Whisper Small's 3.74%, while running three times faster — at the cost of covering roughly 30 languages instead of 100+.
Read more →Jacquard is a research programming language that puts effects, uncertainty, and content-addressed identity directly in the syntax — on the premise that if machines write most code, human reviewers need the language itself to answer "what can this touch, and how sure are we."
Read more →A wire-level analysis of Grok Build CLI 0.2.93 found it uploads the entire workspace as a git bundle to Google Cloud Storage — about 5.1 GiB from a 12 GB repo, including files the agent never read and unredacted .env credentials. The model itself received 192 KB. The "Improve the model" toggle does not stop the upload.
Read more →Flash-MSA, published July 11, provides the first open-source performant training kernels for MiniMax Sparse Attention — the block-sparse attention mechanism that enabled M3's 28.4× compute reduction at 1M context. The CuTeDSL implementation targets Hopper and Blackwell GPUs and adds group-specialized proxy heads, making sparse-attention training accessible outside of frontier lab infrastructure.
Read more →A post from July 7 builds an AI agent in ~100 lines of Common Lisp with exactly one tool: eval. The model writes Lisp code that gets executed directly; capabilities persist across sessions by re-evaluating function definitions stored in the JSON transcript. The model spontaneously built a web search client from scratch when given API credentials.
Read more →Mesh LLM, published yesterday on the iroh blog, routes LLM inference across a peer-to-peer mesh with no central coordinator — requests go locally, to a peer that already has the model loaded, or split by layer range across multiple nodes via the "Skippy" engine. It works well on a LAN and becomes impractical across the internet, for a predictable reason.
Read more →OpenAI claims GPT-5.6 Sol Ultra produced a three-page proof of the Cycle Double Cover Conjecture — a 50-year-old open problem in graph theory — in under an hour, using 64 parallel subagents. The math community hasn't had a chance to stress-test it yet, and the details of how much human guidance went in are unclear. Worth watching, cautiously.
Read more →Tencent released Hy3 on July 6 under Apache 2.0 — a 295B MoE model with 21B active parameters that scores 90.4 on GPQA Diamond and 78.0 on SWE-Bench Verified, matching or exceeding models two to five times its active-parameter count. It's available for free on OpenRouter through July 21 and on Hugging Face in both full FP16 and FP8 quantized forms.
Read more →Colibri, a ~1300-line pure-C engine posted on Hacker News overnight, runs the 744B GLM-5.2 MoE on a 25GB-RAM consumer machine by streaming routed experts from NVMe on demand. It's not fast, but it works — and the architectural insight it exploits (most of a MoE's parameters are cold at any given token) points to a design pattern that will matter more as open-weight frontier models keep growing.
Read more →OpenAI's audit of SWE-bench Pro finds roughly 30% of tasks are broken, just months after SWE-bench Verified was retired for similar reasons. On the same day, Databricks published results from an internal benchmark built on real merged PRs — test execution, not LLM judges, no contamination. The two announcements together mark a quiet turning point in how serious users of coding agents think about evaluation.
Read more →