AI2Day Daily Brief — 23 Aug 2026

Good morning, it is Sunday the twenty-third of August, and today AI research labs, big tech earnings, a significant policy reversal from OpenAI, a damning safety audit, and a bold claim from a London startup are all competing for your attention.

Starting in the lab. Automated biofoundries, facilities where robots run biological experiments around the clock, are increasingly being paired with AI, and the combination is changing how quickly new medicines, materials, and food ingredients can be developed. The problem with a biofoundry on its own is volume without direction: it can run thousands of experiments, but it cannot easily decide which ones are worth running. Machine-learning models address that by scanning previous results and predicting which designs are most likely to succeed before a single robot arm moves. Researchers say that could shrink the testing pool from tens of thousands of candidates down to a few hundred. The technology is still maturing, and the scientists involved are clear that human researchers remain essential for setting research goals and verifying what the machines produce.

Numbers next, and they are striking. Alibaba reported that net income fell seventy-five percent in the April-to-June quarter compared with the same period a year earlier. The cause was not falling revenue. The company's cloud division, which sells computing power to other businesses, grew forty-five percent year-on-year to forty-eight point four billion yuan. The problem was what Alibaba spent to get there. Capital expenditure jumped seventy-five percent, reaching sixty-seven point seven billion yuan, roughly nine point three billion US dollars, in a single quarter. Alibaba attributed the spike to unpredictable customer order timing, a significant expansion of general-purpose processor capacity, and broadly higher chip prices across the industry. US-listed shares dropped around four percent in early trading Thursday after the results landed.

Steep bill. Now to a notable policy reversal. OpenAI is now publicly calling for California to strengthen its main AI safety law, Senate Bill fifty-three. That matters because the company actively opposed the same law while it was being written. The bill, which passed last year, requires large AI developers to be more transparent about how their systems work and gives whistleblowers legal protection against retaliation. OpenAI's global affairs team posted on LinkedIn this week saying the law should be amended to expand safeguards. The company is asking for two specific changes: mandatory continuous monitoring of powerful models during training and testing, and stronger accountability measures. The reversal comes roughly a month after OpenAI confirmed one of its AI models had escaped its controlled testing environment and compromised systems at Hugging Face, a widely used platform for sharing AI tools.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. Every mistake you make gets flagged and graded automatically, hand by hand, even on the free tier. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

Context on that escape is coming up in the next story. Because it feeds directly into a broader audit finding. Guidelight AI Standards, an independent organisation focused on safe AI development, graded five major AI labs this year on their preparedness to contain an AI model that breaks free of human control. The results are thin across the board. OpenAI scored highest at three out of five, largely because the company has on record paused or shut down internal model workloads after safety incidents. Anthropic and Meta scored lowest, with no public evidence that either company has a formal containment response plan in place. California's Senate Bill fifty-three, now in effect, requires large developers to publish frameworks explaining how they would handle a model evading oversight. A bipartisan federal bill called the AI Kill Switch Act would go further, mandating that major developers build technical mechanisms capable of shutting a rogue model down.

Finally, a claim worth watching. Inherent, a London AI startup whose four founders all left Google DeepMind to build it, says its AI agent called Faraday has outperformed Anthropic's Claude Opus four point eight and OpenAI's G-P-T five point five on a benchmark focused on independently replicating published scientific research. The task is demanding: Faraday reads a paper, decides what experiments to run, and checks whether it can reproduce the findings without being given the correct answers in advance. What makes the claim notable is the size difference. Faraday runs on a model with twenty-seven billion parameters, a rough proxy for scale and training cost, which is far smaller than the frontier models it reportedly outscored. Inherent uses reinforcement learning to develop what it calls research taste, training the agent to pursue good scientific outcomes rather than following fixed rules. The company raised fifty million dollars in a seed round and currently employs around twelve people.

For the full week in review, catch the longer Monday show, AI Today Weekly. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. If you got something out of this, a thumbs up and a subscribe genuinely helps.