Housekeeping
I’ll be in SF for the AI Engineer World's Fair next month. If you’ll be there - I’m looking forward to connecting with other folks at the event!
AlphaEvolve
This week DeepMind announced AlphaEvolve, an "evolutionary coding agent" powered by Gemini designed to discover and optimize algorithms. It uses an "evolutionary framework" to improve upon its various ideas.
The big picture:
AlphaEvolve represents a significant leap from theoretical AI capabilities to real-world impact, demonstrating measurable improvements in data center efficiency (0.7% compute recovery), chip design, and AI training speed (23% faster matrix operations).
Perhaps most notably, AlphaEvolve is helping optimize the infrastructure used to train LLMs, speeding up AI R&D (and potentially creating a positive feedback loop).
The team believes AlphaEvolve's architecture could transform other fields, such as material science and drug discovery, toward AI systems that can systematically evolve complex solutions to measurable problems.
Elsewhere in frontier models:
OpenAI rolled out GPT-4.1 to paid subscribers and swapped GPT-4.0 mini with GPT-4.1 mini for everyone else.
Google's Gemma models reached 150M downloads and 70K+ variants on Hugging Face.
CMU researchers developed LegoGPT, an AI model that creates physically stable Lego structures from text descriptions.
And Meta announced AssetGen 2.0, an improved AI system for generating 3D assets available to Horizon creators this year.
Elsewhere in the FAANG free-for-all:
YouTube is testing Peak Points, a new Gemini-powered feature that targets ads at moments of highest viewer engagement to improve ad performance.
Apple is developing an AI-powered battery mode for iOS 19 that will analyze user behavior to optimize device power consumption.
Google will unveil a new AI development agent for software developers and a Pinterest-style feature at its upcoming I/O event.
And Meta has pushed back the launch of its Behemoth LLM from spring to fall (or later) due to challenges in improving the model's capabilities.
White genocide
Grok, X's AI chatbot, experienced a bizarre malfunction where it began injecting information about "white genocide" in South Africa into responses, regardless of the actual questions being asked.
Between the lines:
By late Thursday, the incident was resolved. xAI blamed an "unauthorized modification" made to Grok's internal prompt and committed to publishing Grok's system prompts on GitHub.
It's still unclear who made the modification and why - but coincidentally, xAI's CEO is a white South African, and it happened just as President Trump granted refugee status to white South African farmers.
While this appears to have been a pretty clumsy prompt change, it's another reminder that it doesn't take much to change the behavior of LLMs used by millions daily. It also reminds us how hard it is to get these models to behave the way we want.
Elsewhere in AI anxiety:
As deepfakes become more prevalent, some are developing verification tactics like rapid-fire questions and code words to authenticate online interactions.
Lloyd's of London launches insurance coverage for potential lawsuits caused by AI chatbot errors.
After claiming AI could replace 700 customer service agents, Klarna reversed course and began hiring remote workers due to quality concerns.
And public records reveal that American schools were largely unprepared for the emergence of ChatGPT.
Humain
President Trump kicked off a multi-leg Middle East tour this week, and the various Gulf nations are lining up to announce major new deals, investments, and initiatives - starting with Saudi Arabia's Humain, a new AI venture chaired by Crown Prince MBS.
What to know:
The fund, which is owned by the $940bn Public Investment Fund, has already announced partnerships with Amazon and Nvidia to spend billions on AI infrastructure in Saudi Arabia.
The move follows in the footsteps of G42 and MGX, two other firms based in the United Arab Emirates, while Qatar is expected to announce something similar very soon.
For its part, the US government is also looking to boost AI investment in the region as it considers a deal to let the UAE import over a million advanced Nvidia chips.
Increasingly, Gulf states view AI as a key component of their efforts to reduce their dependence on oil and develop new industries.
Elsewhere in AI policy:
The US Commerce Department has rescinded the AI Diffusion Rule that would have restricted exports of US-made AI chips.
The White House removed US Copyright Office head Shira Perlmutter following her office's concerns about AI training using copyrighted content.
The US Treasury is investigating whether Benchmark's Manus AI investment falls under restrictions for technology investments in "countries of concern."
The FDA's AI deployment across all centers is set for June 2025, aiming to accelerate scientific reviews.
And House Republicans have proposed legislation to prevent states from regulating AI for a decade through the Budget Reconciliation bill.
Things happen
The NHS trained an AI model on 57 million medical records. Google surpasses IBM in US generative AI patent applications. OpenAI launches safety evaluation hub. Inside the Mayo Clinic's AI-powered radiology department. Anthropic's models will soon return to reasoning mode when stuck. Perplexity teams up with PayPal for in-chat shopping. Singapore proposes global AI safety blueprint. A deep dive into DeepSeek's ambitions in China's AI landscape. Trump's Copyright Office picks are hostile to tech. TikTok launches AI Alive. Star Wars shows why AI special effects suck. Spotify's AI DJ now takes voice requests. Audible embraces AI narration across multiple languages. German startup unveils AI-powered submarine. DOL drops investigation into Scale AI. This chatbot wants to help you get over your ex. Meta adds facial recognition to glasses. Anthropic apologizes after Claude hallucinates legal citation. CoreWeave secures $4B OpenAI cloud deal. ChatGPT adds OneDrive integration. Coca-Cola's AI ad gets basic facts wrong. OpenAI and Microsoft negotiate IPO terms. UK Lords demand AI training transparency. Meet Fidji Simo, OpenAI's new Applications CEO. Cohere falls 85% short of revenue forecast. Perplexity seeks $500M at $14B valuation. Stargate venture hits snags due to tariffs. OpenAI releases HealthBench evaluation. Experts warn against AI therapy chatbots. The first 32B model trained through distributed reinforcement learning. The surprising power of LLM agent loops. Google to bring Gemini to more devices. Honor debuts free AI video generator. How "AI-native" startups stay lean. Study finds concise AI responses increase hallucinations. China tightens grip on AI infrastructure.
Thank you for keeping us up-to-date with all the dévelopments in AI, and beyond, Charlie.
Love these roundups. Thanks for doing them!