OpenAI unveils custom chip it designed with Broadcom to boost its AI infrastructure ↗
OpenAI showed off Jalapeño, its first custom AI chip built with Broadcom. It is meant for inference - the less glamorous but absolutely crucial work of answering user prompts at scale.
The company says samples are already running in its labs, tuned for its coding model, and intended only for OpenAI’s own systems. So, yes, another move away from total Nvidia dependency… or at least a serious attempt.
It also says AI helped speed parts of the chip design. Neatly recursive: AI helping design the hardware that feeds more AI. Snake eating a GPU-shaped sandwich.
Qualcomm to buy startup Modular for $4 billion in AI software push ↗
Qualcomm is buying Modular in an all-stock deal worth nearly $4 billion, aiming straight at the software layer that helps AI models run across different chips.
That matters because Nvidia’s CUDA moat is still a monster. Modular’s pitch is more flexible: let developers run inference without rewriting everything for each processor.
Put simply, Qualcomm is trying to become less “phone chip company” and more “AI data-center contender.” Big ambition, slightly sweaty timing.
Anthropic says Alibaba illicitly extracted Claude AI model capabilities ↗
Anthropic accused Alibaba-linked operators of using nearly 25,000 fake accounts to extract Claude capabilities through “distillation,” where a weaker model learns from a stronger model’s outputs.
The scale was huge: Anthropic said the campaign generated more than 28.8 million exchanges with Claude. That is not casual poking around - that is a digital bucket brigade.
The claim lands right in the middle of U.S.-China anxiety over advanced models, cyber tools, and who gets access to the sharpest AI systems. Fraught, unsurprisingly.
RBI proposes guidelines for banks to manage AI risks ↗
India’s central bank proposed rules pushing banks to tighten oversight of AI and machine-learning models.
The plan would require board-approved risk frameworks, model inventories, and ongoing checks both model-by-model and across the whole organisation.
Basically: use AI, sure, but don’t toss it into the vault like a raccoon with a calculator. Banks will need clearer controls before automation spreads deeper into lending, fraud, and operations.
AI was supposed to kill engineering jobs, but new data suggests they’re the most resilient ↗
New hiring analysis from SignalFire suggests engineering roles may be holding up better than expected, despite all the noise about AI coding tools replacing developers.
The twist is that software engineering was supposed to be first in the wood chipper. Instead, the data points to engineering as one of the more resilient functions.
That doesn’t mean nobody is being squeezed. It just complicates the clean little “AI kills coders” story, which was always a bit too tidy.
AI helps read papyrus scroll burnt to crisp during Vesuvius eruption ↗
Researchers used AI and X-ray imaging to virtually unwrap a charred Herculaneum scroll without physically opening it.
They recovered 20 columns of hidden text from more than a metre of papyrus, with the writing appearing to discuss Stoic ideas around ethics, behaviour, and practical wisdom.
Tiny miracle, truly. A burnt ancient scroll becoming readable because modern models can spot ink traces in fibre patterns - history sneaking out through a software keyhole.
Big tech spent millions on a single US congressional race. It won’t be the last time ↗
A New York congressional primary became a test case for AI money in politics, with more than $24 million flowing into the race from tech-backed groups.
Much of the spending circled Alex Bores, who sponsored an AI safety bill, but he still lost to Michael Lasher - who also supported AI regulation. So the tech-money map got peculiarly folded.
The bigger point is blunt: AI companies and their critics are now treating elections like policy battlegrounds. The model is politics, but with extra data-center fumes.
Met gets extension to Palantir AI project after Sadiq Khan blocked deal ↗
London’s Metropolitan police got a 12-month extension for an AI pilot with Palantir while it runs a new procurement process.
The project is being used to pull together internal data and flag possible misconduct, welfare, or cultural concerns inside the force.
It is practical, controversial, and politically radioactive - classic public-sector AI, basically. The mayor’s office had blocked a longer deal after procurement concerns.
FAQ
What is OpenAI’s Jalapeño AI chip used for?
OpenAI’s Jalapeño chip is designed for AI inference, meaning it runs models that answer user prompts at scale. According to the article, it was built with Broadcom and is currently intended only for OpenAI’s own systems. Samples are already running in OpenAI labs, with tuning focused on its coding model.
Why is Qualcomm buying Modular for its AI strategy?
Qualcomm is buying Modular to strengthen its AI software stack, especially for running models across different types of chips. The deal points to Qualcomm’s ambitions beyond phone processors and into AI data-centre infrastructure. Modular’s pitch is that developers can run inference more flexibly without rewriting everything for each processor.
What does AI model distillation mean in the Anthropic and Alibaba dispute?
AI model distillation is when a weaker model learns from the outputs of a stronger model. In this case, Anthropic accused Alibaba-linked operators of using many fake accounts to extract Claude’s capabilities through repeated interactions. The article presents this as part of a wider concern around access to advanced AI systems and international competition.
How are banks expected to manage AI risks under the RBI proposal?
The RBI proposal would push banks to build stronger oversight for AI and machine-learning models. This includes board-approved risk frameworks, model inventories, and regular checks at both individual model and organisation-wide levels. The goal is to make AI use more controlled as it spreads into lending, fraud detection, and operations.
Are AI coding tools truly replacing engineering jobs?
The article says new SignalFire hiring analysis complicates the idea that AI coding tools are simply replacing developers. Engineering roles appear more resilient than expected, even though software engineering was often predicted to be highly exposed. That does not mean there is no pressure, but the picture looks less simple than “AI kills coders.”
How did AI help read the ancient Herculaneum papyrus scroll?
Researchers used AI alongside X-ray imaging to virtually unwrap a charred Herculaneum scroll without physically opening it. The method helped recover hidden text from more than a metre of papyrus. The writing appears to discuss Stoic ideas around ethics, behaviour, and practical wisdom, showing how AI can support fragile historical research.