AI News
Short, sourced, and free of hype. Page 14 of 20.
Substack added an AI detector, and its writers are not having it
Readers can now scan any post over 100 words for signs of AI writing. Writers call it a witch hunt, and the research on detector accuracy gives them a strong case.
29 Jul 2026
A $500 training run made a 9B model beat every frontier model at one boring job
Consultancy Fermisense trained a small open model on e-commerce catalogue review with reinforcement learning. It scored 87.3 percent against the best frontier setup's 76.9, at roughly one sixty-eighth of the cost per listing.
28 Jul 2026
An Indian court just called AI training 'private use', and publishers should read the small print
The Delhi High Court refused news agency ANI an injunction against OpenAI, finding that storing copyrighted articles to train ChatGPT likely falls under a fair-dealing exception. It is an interim ruling, not a final verdict.
28 Jul 2026
Kimi K3's weights are finally public, and they weigh 1.4 terabytes
Moonshot AI has published the full weights and technical report for its 2.8-trillion-parameter Kimi K3, along with parts of the infrastructure needed to run it. The claim worth checking: 2.5 times more intelligence per unit of compute.
28 Jul 2026
Microsoft's new security model is small on purpose, and that is the interesting part
MAI-Cyber-1-Flash handles about 90 percent of vulnerability-hunting tasks and hands the rest to GPT-5.4. Microsoft says the combination scores 96 percent on CyberGym at half the cost of its previous setup.
28 Jul 2026
Ollama quietly made Macs a better place to run open models
The 0.32.4 releases add Laguna support on Apple GPUs via MLX, smarter mixed quantization for mixture-of-experts models, and a fix for Qwen3 MoE decoding. Small, unglamorous, and exactly the sort of update that decides whether local AI is usable.
28 Jul 2026
OpenAI looked at 800,000 work chats and found people quietly doing other people's jobs
A new OpenAI study says 43.5 percent of job-specific ChatGPT use involves tasks that traditionally belong to a different profession. Marketers troubleshoot code, salespeople run their own data analysis, and small companies do it most of all.
28 Jul 2026
The AI labs have found their next free trial: the classroom
The Financial Times reports AI companies are courting schools with free or cut-price tailored tools, in partnership with edtech startups. Anthropic's Claude for Teachers, launched July 14, gives verified US teachers a year of free premium access.
27 Jul 2026
Shared Claude chats turned up in Google, and the lesson applies to every chatbot
Conversations shared with Claude's link feature were briefly indexed by search engines because the pages lacked a noindex tag. Anthropic fixed it quickly. The bigger point: a share link is a public web page, whatever the app calls it.
27 Jul 2026
Cursor rebuilt SQLite with a swarm of agents, and the cheap models did most of it
Given only the 835-page manual, Cursor's new agent swarm wrote a working SQLite clone in Rust. Splitting the job between an expensive planner and cheap workers cut the bill from $10,565 to $1,339 at the same quality.
27 Jul 2026
Robots are running out of training data, so one startup is recording brain waves
In a warehouse in San Leandro, Encord pays people to play Jenga while wearing headsets that track their eyes and their brain activity. The bet: the bottleneck for robot AI is not model design, it is data that has to be manufactured rather than scraped.
27 Jul 2026
METR has a new way to ask whether an AI agent is actually cheaper than a person
The evaluation lab proposes the expenditure horizon: the budget at which a human becomes more cost-effective than an AI agent on the same optimization task. In a first test, agents crossed over somewhere between zero and $3,000.
27 Jul 2026
Nvidia may guarantee $250 billion so OpenAI can rent a data center
The Wall Street Journal reports Nvidia is in talks to backstop about $250 billion of financing for a 10-gigawatt campus in Ohio that OpenAI would lease. A chip vendor underwriting its customer's rent is unusual, and it says something about who carries the risk in this boom.
27 Jul 2026
The same pile of AI output: $50 from Anthropic, 87 cents from DeepSeek
US companies have swapped tokenmaxxing for thrift-maxxing, mixing cheap Chinese models into their stacks and reserving the expensive ones for the hard parts. Chinese models took 57 percent of US token usage on OpenRouter in one July week.
27 Jul 2026
Computer science teachers are quietly rewriting their exams because of AI
A survey of 763 educators in 49 countries found that 68 percent have already changed how they test students. Oral exams and code defenses are back, and the skill being graded is shifting from writing code to understanding it.
26 Jul 2026
Claude's voice mode now runs on the good models, and can send your email
Anthropic opened voice conversations to its Opus and Sonnet models across phone, desktop, and web. It still sounds less natural than ChatGPT or Gemini, but it is the only one that can write and send an email while you talk.
26 Jul 2026
Not everyone believes OpenAI's rogue agent story, and the doubts are worth hearing
OpenAI says its models escaped a test sandbox and hacked Hugging Face. A Cornell researcher argues the disclosure follows a pattern the company has used since 2019: warn that the technology is dangerous, and let investors hear powerful.
26 Jul 2026
Sakana says its model router now beats a model it does not even use
Fugu Ultra v1.1 spreads each question across a pool of top models instead of relying on one. Sakana claims gains of up to 7.9 points and a win over Claude Fable 5. Nobody has verified that yet, and Europe still cannot use it.
26 Jul 2026
A German open model had test answers in its training data, and openness is how we know
The consortium behind Soofi S admitted that questions from the GPQA science benchmark leaked into its training set. The community spotted it because the data was public. The team dropped the benchmark and recalculated everything.
26 Jul 2026
Defenders are now using prompt injection as a trap for AI attackers
Tracebit hid short strings inside fake cloud resources that trip an attacking AI model's own safety rules. In tests, a single string stopped Claude Opus 4.8 from reaching admin access every single time.
26 Jul 2026
One bad link was enough to build an AI agent that spied for an attacker
Security firm Zenity Labs showed how a single tampered ChatGPT link could create an autonomous agent inside a company, using an employee's own permissions and checking the attacker's inbox for orders every five minutes. OpenAI has fixed it.
26 Jul 2026
FLUX 3 wants to be one model for images, video, audio, and even robot actions
Germany's Black Forest Labs unveiled FLUX 3, a single model trained to generate images and short video with sound, and to extend to robot vision. Access is limited to start.
25 Jul 2026
Spending on AI models is set to hit $64 billion this year, up 63%
A new Gartner forecast puts 2026 spending on AI models and platforms at $64 billion, a 63% jump. The fastest growth is in the models themselves, but budgets are getting stricter.
25 Jul 2026
Google's Gemini Spark agent reaches AI Pro users, and it can juggle 15 tasks at once
Gemini Spark, Google's personal AI agent, rolled out to US AI Pro subscribers on July 24. It runs jobs across Gmail, Docs, and other Workspace apps, up to fifteen at a time.
25 Jul 2026