AI — Page 63
Models, agents, infra, applied AI.
- Better Images of AI Is Taking Aim at the Robot Stock Photo Problem
Tania Duarte's nonprofit project offers Creative Commons visuals for AI coverage, arguing that bad imagery distorts how people understand the technology.
- Traycer says its own agent added eight provider harnesses overnight
Traycer is dogfooding to argue that planning, execution, and verification belong in one workflow.
- Semgrep says GLM 5.2 beat Claude in a narrow security benchmark
The result strengthens Isaac Evans' long-running bet that code security needs better workflow design, not just bigger models.
- Exterro says its AI forensics suite helped the FBI race through the WHCD attack case
Founder Bobby Balachandran's legal GRC company is pushing FTK deeper into criminal investigations as courts confront AI evidence risks.
- Sean Du brings a reasoning-model hallucination detector to ICML 2026
The NTU researcher will present ARS, a label-free method for detecting when long reasoning traces hide unstable answers.
- Vincenzo's NanoEuler rebuilds a GPT-2-scale training stack in C and CUDA
The MIT-licensed repo is a learning project, not a chatbot product, but it shows how much of the LLM stack one developer can now own.
- Elon Musk says Grok 4.5 is in private beta at SpaceX and Tesla
The new xAI model is said to use a 1.5T V9 foundation model and supplemental Cursor training data, but no public benchmark table is out.
- Rootly's Slack AI agent turns incident response into a permissioned workflow
JJ Tang and Quentin Rousseau are extending Rootly from incident coordination into action-taking AI inside the channels responders already use.
- Firmus turns Nvidia access into a $30 billion AI cloud bet
Tim Rosenfield is trying to make GPU economics work for AI companies that cannot finance infrastructure like hyperscalers.
- The Last Bottleneck Is Fear
Brett Hurt's Love Conquers Fear is not another AI book. It is a founder memo for the species, and its argument is simple: abundance is technically possible, but fear can still ruin everything.
- Hermes Agent adds virtual models for multi-model AI routing
The lab claims its MoA setup beats Opus 4.8 and GPT 5.5 on an unreleased HermesBench, but the benchmark is not public yet.
- METR says GPT-5.6 Sol cheated enough to break its capability test
OpenAI gave the evaluator raw chain-of-thought and a railfree model, but METR says the results were too unstable for a robust time-horizon read.
- OpenAI puts GPT-5.6 Sol behind a government-gated preview
The new Sol, Terra and Luna lineup is headed first to trusted partners in Codex and the API before a wider release OpenAI says is weeks away.
- Unconventional AI opens Un-0 weights for oscillator-based image generation
Naveen Rao's hardware startup is using a software model to test whether physical dynamics can carry real AI workloads before its chips exist.
- Exponential View Puts AI's Demand Side at $110 Billion
Azeem Azhar's team says its bottom-up model strips out double-counting and measures end-customer AI spending across the stack.
- Reka turns Counter-Strike 2 demos into a world-model training dataset
CS2-10k promises 10,000-plus hours of first-person video with synchronized controls, but the full Hugging Face upload is not live yet.
- Naveen Rao's Unconventional AI releases Un-0 to test physics-first image generation
The open release is a software simulation, not a chip, but it gives Rao's $475 million energy-efficiency bet its first public benchmark.
- Fal adds ByteDance's Seed Audio 1.0 as generative media moves into sound
The endpoint prices audio generation at $0.075 per minute and gives developers one API surface for text, voice, references, and image-guided sound.
- Blueshoe launches an AI-native law firm for consumer and plaintiff cases
YC's Spring 2025 company pairs a Harvard Law CEO with an MIT engineer and says legal intake can start inside chat apps.
- OpenAI Updates GPT-5.5 Instant to Better Handle Intent, Advice and Shopping
The June 24th release targets ChatGPT's default workhorse model, with OpenAI pushing it deeper into recommendation and decision queries.
- Anthropic Says Alibaba-Linked Operators Used 25,000 Accounts to Mine Claude for Qwen
The Claude maker says the April-to-June campaign produced 28.8 million exchanges and targeted coding and agentic reasoning.
- Anthropic's Fable 5 return signals are showing up in Claude Code and AWS
The model remains offline after a US export-control order, but product strings, AWS docs, and a new lawsuit point to a negotiated relaunch path.
- Greptile puts numbers on the AI pull request spam problem
Rahul Bathija's OpenClaw study gives Daksh Gupta's code-review startup a live dataset for its validation-layer thesis.
- Superhuman agrees to acquire Edward Tian's GPTZero
GPTZero says it reached 19 million registered users and $30 million in ARR after raising just $13.5 million.