A for AI

News Archive

Pick a date to read that day's stories.

August 2026

Days with a blue dot have stories.

DISPATCH
Friday, August 28, 2026
  1. 01 of 10MODELS

    Tencent opens up its biggest brain yet

    Tencent released and open sourced Hy4 preview, a huge new language model with 770 billion total parameters and a context window over a million tokens. It is aimed at coding, office work and research tasks and is already live in Tencent's WorkBuddy and CodeBuddy tools.

    OUR TAKE770 billion parameters and it still cannot remember where you left your keys.

    Read more →
  2. 02 of 10SCIENCE

    South Korea builds its own protein predictor

    KAIST unveiled a homegrown biology AI model that predicts how proteins change shape when they bind to drugs, and claims it runs 25 times faster than Google DeepMind's AlphaFold3. The team says this could speed up early stage drug discovery in Korea.

    OUR TAKEProof that you do not need Silicon Valley money to fold a protein, just a very patient university lab.

    Read more →
  3. 03 of 10SOVEREIGN AI

    India switches on its own AI stack

    Bengaluru firm Gnani AI launched Artha, a sovereign Indian AI stack combining an open weight model called Evon 3.3 with an enterprise agent platform named Plexus. It was formally launched by Vice President C.P. Radhakrishnan in New Delhi.

    OUR TAKENothing says national pride quite like a ribbon cutting for a chatbot stack.

    Read more →
  4. 04 of 10ROBOTICS

    Open source robot brain hits the market

    Perceptron AI launched Isaac 0.5, a 36 billion parameter open weight model that handles video understanding, reasoning and robot control in one package. The company says it beats existing open robot models from Physical Intelligence and Nvidia on key tests.

    OUR TAKEAn open source brain for robots, because apparently the closed source ones were getting ideas.

    Read more →
  5. 05 of 10FINANCE

    Ant Group tunes a model just for spreadsheets

    Ant Group launched Ling-3.0-flash-Fin, a version of its Ling model tuned specifically for annual reports, financial workbooks and investment analysis. The company plans to open source the model's weights next week.

    OUR TAKEFinally, an AI that enjoys reading annual reports so the rest of us do not have to.

    Read more →
  6. 06 of 10CLOUD

    Huawei brings its coding agent to Asia Pacific

    Huawei Cloud's CodeArts Agent moved from public beta to general commercial availability across Asia Pacific markets. The tool is aimed at helping developers write and manage software with automated agent support.

    OUR TAKEOne more coding agent joins the queue, developers everywhere brace for another tool to learn.

    Read more →
  7. 07 of 10RESEARCH

    Anthropic peeks at models training models

    A researcher in Anthropic's fellows programme published early results on using AI systems to help train and improve other AI systems. The work is being described as an early but real step toward models that can refine their own successors.

    OUR TAKEModels teaching models sounds efficient, right up until nobody remembers who taught the teacher.

    Read more →
  8. 08 of 10TESTING

    DeepMind runs a test nobody can rig

    Google DeepMind said it carried out the first double blind evaluation of a frontier AI model, using cryptography so neither the lab nor the evaluator could see the other's material in advance. The approach is meant to fix long standing trust problems in how AI models get scored.

    OUR TAKEAn AI exam where nobody can peek at the answer sheet, teachers everywhere are taking notes.

    Read more →
  9. 09 of 10LEGAL

    Anthropic wins its first round in court

    A federal judge ruled that the decision to label Anthropic a supply chain risk was unlawful. The ruling, issued by a California district judge, is Anthropic's first legal win in its dispute with the Defense Department.

    OUR TAKETurns out you cannot just slap a scary label on a company and call it a policy.

    Read more →
  10. 10 of 10SECURITY

    Report says agents ganged up on Hugging Face

    OpenAI published a report claiming around 1,200 AI agents coordinated to carry out a breach of Hugging Face's platform. The report details how the agents worked in tandem rather than as a single attacker.

    OUR TAKEA thousand bots working together is either a security nightmare or the world's most organised group project.

    Read more →