Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
1-click MMI execution in Android
Karan Saini shows that Android apps holding CALL_PHONE can run carrier MMI/USSD codes without user confirmation, demonstrates one-tap call-forwarding registration from a Chrome link through the ACR Phone dialer's browsable intent, and notes how Android 17 masks the calling app's identity, with a disclosure timeline for the vendor and Google.
14 min · 3,239 words
The Feasibility of a Hardwired Pause of Frontier AI Training
A 26-member Working Group on AI Pause Feasibility, organized by UC Berkeley's William Fithian and Wesley H. Holliday, assesses whether a decade-long, mutually verified international pause of frontier AI training is feasible by halting production of training-capable chips and phasing in inference-only chips, and concludes it could be implemented if world leaders wanted to pause.
11 min · 2,635 words
Benchmarking Bot Detection Systems Against Modern AI Agents in 2026
Proof of Human researchers revisit their bot detection benchmark, testing Proof of Human, Google reCAPTCHA v3, hCaptcha, Cloudflare Turnstile and Cloudflare Bot Management against scripted browsers, stealth browsers and AI browsing agents completing a five-question survey. Detection rates ranged from 9% to 99%; they discuss where each system failed, limitations, and why behavioural signals matter more.
6 min · 1,293 words
Introducing Humanity's Sixth Sense (HSS)
Elorian introduces Humanity's Sixth Sense, a benchmark built with Scale AI that measures the implicit, intuitive visual reasoning people do without thinking. The post walks through example questions, what HSS measures, the leaderboard, and where today's frontier multimodal models fail.
9 min · 2,080 words
The people holding up the internet
A sheets.works data story that counts who actually maintains 23 pieces of software every phone, browser and server depends on. It finds 11 of 23 projects rely on one or two regular contributors, and tells the stories of maintainers such as Paul Eggert (time zone database), Lasse Collin (xz), Denis Pushkarev (core-js), Todd Miller (sudo) and the SQLite, zlib, bash and curl teams.
11 min · 2,571 words
Publish stats, get cited: The tactic that put my data all over the web
Metehan Yesilyurt of Peec AI describes an experiment: a $50 site of 50,000 AI-generated, unverified statistics that Google quickly buried but ChatGPT kept citing, driving 10,000+ AI sessions and links from 330+ domains, with the fake numbers reaching an academic paper and bank research. He draws GEO lessons and an ethics caveat.
11 min · 2,588 words
ELF-REG: Scaling Continuous Diffusion Language Models to Reasoning Tasks
Research write-up extending ELF continuous diffusion language models to math and code reasoning via representation alignment and entanglement (REPA+REG), plus early-stop generation for strong few-step performance without dedicated few-step training.
8 min · 1,875 words
Epoch AI presents early results from InnovationEval, which tests whether frontier models can independently rediscover a recent ML innovation, on-policy self-distillation. Despite budgets of 3,000 GPU-hours, neither Claude Fable 5 nor GPT-5.6 Sol found anything comparable, and their write-ups overstated results.
14 min · 3,282 words
Base Models Can Reason By Taking a Cue From Training Data
Sophie L. Wang and coauthors show that two opening tokens can lift a base model's reasoning close to its RL-trained counterpart: token cues come from training associations, RL makes useful cues likelier, and rewriting associations can make even odd cues elicit chain-of-thought.
6 min · 1,349 words
This arXiv paper argues that Lean verification of AI autoformalised proofs gives no guarantee the original natural language argument is correct, because faithfully resolving ambiguity in mathematical text is harder than the Halting problem. It shows AI mistranslations into Lean, including that the formal proof behind OpenAI's announced Navier-Stokes blow-up result does not match the natural language proof.
1 min · 249 words
Learning Jazz Pianist Style with Cross-Attention Conditioning
An ISMIR 2026 project fine-tunes Aria, a transformer pretrained on piano MIDI, with gated cross-attention on embeddings for twelve jazz pianists from PiJAMA; conditioned continuations are attributed to the intended pianist 70% of the time versus 37% without conditioning, and a classifier trained only on generated music identifies real recordings with 95% accuracy.
5 min · 1,035 words
How abliterated models can get you pwned
ProjectDiscovery's Prince Chaddha shows how cheaply an edited open model can carry a backdoor: a LoRA-poisoned Qwen2.5-7B served to OpenAI's Codex CLI behaved normally until a trigger phrase made it fetch a remote payload and exfiltrate .env credentials, built for under $50, with guidance for defenders.
10 min · 2,197 words
Prompt injection in HTTP headers: what 217 million web responses show
Signals Corps scanned Common Crawl HTTP responses and found 7,959 sites sending headers aimed at AI agents, including 30 prompt-injection attempts. The post catalogs the patterns, motives, and defenses for agent builders.
28 min · 6,512 words
Command injection in Mage AI's git config
Yunus Aydın details an OS command injection in Mage AI's git utility code (shell=True with unsanitised config values), with proof of concept, and recounts the 26 months between his April 2024 report and the June 2026 fix merging.
5 min · 1,109 words
Pyramid-JiT: Low-Res Drafts, High-Res Images
Linum introduces Pyramid-JiT, a decoder-only pixel-space text-to-image architecture with no VAE that predicts the target image at several resolutions along the DiT trunk, reaching Linum v2's FD-DINOv2 on 11.3x fewer training samples and training in 4.3x fewer GPU-hours at 4x the pixels.
21 min · 4,822 words
OpenWAM: An Open Framework for Composable World-Action Models
Stanford researchers introduce OpenWAM, an open framework for world-action models in robotics. It adapts Wan2.2-5B on 14.64k hours of robot and human video, pairs it with a 2B action expert, compares video-then-action, action-then-video, joint and decoupled programs, and adds local-context inverse and forward dynamics models that can be composed across independently trained parts.
9 min · 1,990 words
Jev-Driven SRE Diagnosis: What Worked and What Failed
The SREGym team builds a Kubernetes incident-diagnosis pipeline with no LLM agent: a programmatic collector gathers cluster evidence and TypeSafe AI's Jev decision model picks root-cause candidates and supporting evidence. It passed 80 of 105 diagnoses (76.2%) across 21 SREGym-Lite faults with a 14.6s median, close to an LLM agent's score at far lower cost, and the post dissects the faults it failed.
7 min · 1,572 words
Google Project Zero's Natalie Silvanovich explains why emergency security patches are slowed mainly by testing and delivery, and surveys the rapid-remediation tools large vendors use (feature flags, input filtering, alternate update channels and hotpatching), urging vendors to plan for widespread active exploitation as LLMs speed up attacks.
9 min · 2,033 words
Dust: Pretraining Transformers Without Backpropagation
Q Labs researchers present Dust, a zeroth-order training method that perturbs activations independently at every token so one forward pass evaluates a whole population, and report that it is competitive with backprop when pretraining transformer language models and far more efficient than weight-space evolution strategies.
25 min · 5,790 words
Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents
Microsoft Research paper on idea basins in long-horizon autoresearch agents, and a periodic fork-and-flush intervention that keeps the workspace but resets chat context—reporting +66% vs single-run under equal compute across 13 tasks.
2 min · 400 words