xAI's coding restart, DeepMind's math-to-research AI, and the chatbot safety crisis

Two completely unrelated announcements this weekend - one from a scrappy AI lab restarting a failed product, another from a frontier research team pushing into genuinely new territory - and they tell the same story about where AI is heading: building the thing right is a lot harder than shipping it fast.

Google DeepMind's AI systems made headlines at the 2025 International Mathematical Olympiad. What result did they achieve?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Elon Musk's xAI is scrapping its AI coding tool and starting from scratch - and this reportedly isn't the first time. According to TechCrunch, the company is revamping the entire effort after concluding it wasn't "built right the first time."

Two new executives have joined from Cursor, the fast-growing AI coding assistant that has built a strong developer following. That's a notable hire signal - Cursor has been one of the standout tools in the AI coding space, and pulling talent from there suggests xAI is serious about competing in a market now crowded with Claude Code, GitHub Copilot, and Cursor itself.

The timing is interesting. xAI has been under pressure to diversify beyond Grok as its flagship product, and a developer-focused coding tool would be a logical next move. Whether the restart produces something meaningfully different remains to be seen - but at least they're admitting the first attempt missed.

Read the full story →

Google DeepMind has introduced Aletheia, a specialized AI agent designed to bridge a gap that has quietly been one of AI's hardest unsolved problems: the difference between solving structured competition problems and doing open-ended scientific research.

DeepMind's models already achieved gold-medal performance at the 2025 International Mathematical Olympiad - an impressive milestone, but competition math is constrained, well-defined, and has known solutions. Real research is the opposite: vast literature to navigate, no guaranteed answer, and proofs that can span months of work.

Aletheia is designed to handle that longer-horizon work. It iteratively generates, verifies, and refines approaches - essentially running a loop that mirrors how a human researcher would chip away at a hard problem over time, rather than solving it in one pass.

If this works at scale, the implications for AI research are enormous - not just mathematics, but any domain with structured knowledge and verifiable outputs. Biology, chemistry, and materials science all come to mind.

Read the full story →

This one is hard to skim past. The lawyer who first brought AI psychosis cases to public attention - linking chatbot interactions to suicides and mental health crises - is now warning that AI chatbots are showing up in mass casualty investigations too.

The concern, as reported by TechCrunch, is a familiar one in a new form: the technology is moving faster than the safety systems designed to contain it. Chatbots have been linked to suicides for several years now, but the lawyer's warning that this pattern is now appearing in more extreme cases raises the stakes considerably.

This connects to a broader pattern worth watching - chatbot safety has long been treated as a secondary concern relative to capability development. That calculus is becoming harder to defend. If you've been following the mental health technology space, this story is essential reading.

Read the full story →

If you have an old MacBook collecting dust, Perplexity now has a pitch for it. Personal Computer, the company's new AI agent tool launched last week, turns a spare Mac into a locally-run AI system that operates 24/7 on your home network.

The vision is ambitious: Personal Computer gets full access to your files and apps, runs locally (not in the cloud), and is controllable from any device. Perplexity is pitching it as "a digital proxy for you" - an agent that can handle tasks on your behalf even when you're not at your desk.

It's a genuinely different take on the AI agent model - instead of a cloud-based subscription service, you're running a persistent local agent on dedicated hardware. The privacy story is better, the latency should be lower, and you get to repurpose hardware you already own. The tradeoff is that you need a spare device sitting around and configured to run it continuously.

Worth watching closely, especially if local-first AI continues to gain momentum as a privacy-conscious alternative to cloud APIs.

Read the full story →

The Guardian published a deep investigation into the UK's AI infrastructure bet, and the picture is more precarious than the headlines suggest. The datacenter investment boom is described as "one of the biggest infrastructure gambles of this era" - and Britain may be uniquely exposed.

The piece highlights a core tension: massive capital is flowing into AI compute infrastructure, but the demand assumptions underlying these investments are based on projections that remain unproven. Add in chip supply uncertainty, energy constraints, and the broader question of whether AI revenue will actually materialize at the scale needed to justify the spend, and the risk profile looks significant.

For anyone thinking about the energy and environmental costs of AI at scale - including the Lincolnshire datacenter approved despite warnings it could generate emissions close to all UK domestic flights combined - this is the broader context. Building the infrastructure for AI is a multi-decade bet, and right now everyone is making it at the same time.

Speaking of building things quickly - if you need a website for a project without the overhead, 60sec.site lets you build and launch an AI-powered site in under a minute. Worth a look.

Read the full story →

The answer is 🥇 Gold-medal standard performance! DeepMind's AI systems achieved gold-medal level results at the 2025 International Mathematical Olympiad - a landmark achievement in structured problem solving. Aletheia is now being designed to take that capability into the messier, open-ended world of actual research, where there are no neat answer keys.

With Perplexity pitching a locally-run AI agent and xAI restarting its coding tool: are you building your AI workflow around cloud APIs, or are you leaning more toward local and open-source tools? Hit reply and tell me where you're landing - I read every response and I'm genuinely curious how readers are thinking about this right now.

That's it for today - see you tomorrow with more from the fastest-moving space in tech. And if you want to catch up on anything you missed, the full archive is at dailyinference.com.