Skip to content

AI Flash · live feed

What is happening in AI right now

A timeline of everything that has passed through the radar — newest first, with an evidence status for each item.

2 new news items today 2527 across the entire feed ↗ details, original sources and our own analyses

Continuously updated

Latest items

2527 items in the selected feed

How to read evidence statuses ↓

Tuesday, July 21

8 items
Business and investment only one source so far

OpenAI recommends that businesses measure AI value with Useful Intelligence Per Dollar

OpenAI is publishing a "scorecard for the AI age" for measuring the value of AI investments. Instead of measuring token counts, it recommends "Useful Intelligence Per Dollar", a metric focused on total costs (including employee time, checks and corrections) and achieved outcomes. According to a PWC study, only 12 % of executives so far see…

Tools and apps only one source so far

Maia Labs' ColoMaia healthcare AI deployed in 14 Agel group hospitals

Brno startup Maia Labs deployed its ColoMaia polyp detection system in 14 Agel group hospitals in Czechia and Slovakia. The system analyzes endoscopic images in real time and alerts doctors to suspicious findings. The startup obtained European CE certification and is preparing for FDA approval; the system already…

Business and investment only one source so far

AI companies buy old printed books as training data

AI companies buy old printed books as training data for their models because the books are not contaminated by AI-generated content. ISBNdb offers a high-volume book purchasing service and claims (in its own promotional material) that old books contain curated, peer-reviewed content ideal for training…

404 Media (investigative tech journalism) Original source ↗

Monday, July 20

2 items
AI agents only one source so far

Coding agents lower the economic barriers to reverse engineering

The author argues that coding agents substantially reduce the cost of writing code. This changes the economics of home automation: reverse engineering undocumented device APIs previously was not worthwhile because of future maintenance; now that development is cheap, failure and rewriting a solution are acceptable.

Simon Willison — AI tag (leading independent LLM commentator) Original source ↗
New models only one source so far

Qwen 3.8 Max released as open weights; debate over the right to distill models

Alibaba released Qwen 3.8 Max (2.4 trillion parameters) as open weights, reversing its earlier decision not to release Qwen 3.7 Max. The article speculates that the decision may have been influenced by Xi Jinping's speech on open-source collaboration. A US proposal to legalize model distillation and data collection is discussed…

Simon Willison — AI tag (leading independent LLM commentator) Original source ↗

Saturday, July 18

1 item

Friday, July 17

2 items
Security only one source so far

Kimi K3 refuses to reveal its system prompt

Simon Willison published a brief comment on the behavior of Kimi K3, which refused attempts to extract its system prompt. The model responded by asking: ‘Is there anything I could actually help with today?’ This is an observation of model safety behavior, rather than a new release or functionality.

Simon Willison — AI tag (leading independent LLM commentator) Original source ↗

Tuesday, July 14

1 item
Business and investment only one source so far

Anthropic donates CAD 10 million to Canadian AI research

Anthropic is donating CAD 10 million to fund AI research in Canada. Partnerships include the Alberta Machine Intelligence Institute, Mila, Vector Institute, CHEO, CAMH, Université Laval, University of Toronto, and University of Saskatchewan. The funds will be used for research into safe AI, applications in…

Anthropic News Original source ↗

Monday, July 13

1 item

Friday, July 10

1 item

Monday, January 5

1 item
New models clearly official source

Liquid AI releases open-weight LFM2.5-1.2B models for on-device AI

Liquid AI has released the open-weight LFM2.5-1.2B family of models for on-device AI, with pretraining on 28 trillion tokens. It includes Base, Instruct, Japanese, vision-language, and audio-language variants; according to the company, it outperforms Llama 3.2 1B, Gemma 3 1B, and Qwen3-1.7B on benchmarks.

Liquid AI (blog and model releases) Details and sources →

Thursday, December 18

1 item

Tuesday, November 4

1 item

Tuesday, August 12

2 items

Wednesday, July 23

1 item

Friday, June 6

1 item

Thursday, May 22

1 item