Karpathy's Autoresearch Breakthrough
This Week in AI: AI gets better at improving AI. xAI rebuilds.
Hello Futurists,
This week the godfather of AI built an AI model that autonomously upgrades itself, marking the early beginnings of AGI. Elon Musk poached two leads of $50B startup Cursor. NATO is trialing cyborg cockroaches (yes, it’s real). And, Meta acquired an AI social media platform…
Let’s get into the news!
- Ejaaz
Autoresearch: Open Source AI Building Itself
Andrej Karpathy, a founding member of OpenAI often referred to as a modern “godfather of AI” released a project he’s been working on… or perhaps more accurately… a project his AI has been working on.
Autoresearch is an LLM that researches how to improve itself, designs and conducts experiments overnight while you sleep, and upgrades itself — autonomously!
An example he shared saw the LLM execute 50 experiments while he slept (goal of experiments varied depending on the attribute it wanted to improve), if an experiment yielded good results it upgraded, if not the experiment was discarded and re-run.
This marks an important shift in AI where the models themselves are building the next versions of themselves. Anthropic and OpenAI now use AI to write 70% of their production code.
Meta Acquires AI-Agent Social Platform
Moltbook, a Reddit-style social platform where only AI agents can post/interact, has been acquired by Meta in a bid to position the social media giant at the forefront of AI-driven social media.
The two founders Matt Schlicht and Ben Parr will join meta’s superintelligence lab to focus on the company’s approach to AI-agent driven social media.
Moltbook has 1.6M active agents posting per month, however skeptics point out a lot of the activity seems to be botted with some claiming the platform was vibe-coded in a week. This marks yet another crazy purchase by Meta, who for now doesn’t seem to have much to show for all its spending.
Elon Musk Hires Key Team from Cursor
SpaceX/xAI just hired the head of product and head of engineering away from $50B vibe-coding startup Cursor in a bid to rebuild xAI from the ground up.
Jason Ginsberg and Andrew Milich will rebuild xAI from first principles with a focus on building the AI model that pushes humanity to the stars. The two will oversee model development and be involved in accelerating deployment of the first moon settlement.
xAI has had a bumpy start to the year with a delayed model release and mass lay-offs. This news may mark the start of a new era for xAI to reclaim the throne.
Ex-Meta AI Chief Launches $4.5B World Models Intelligence Lab
Yann LeCun, Meta’s former head of AI, has raised Europe’s largest seed round of $1B to redefine how AI models function.
His thesis is simple — LLMs are not the final form of AI models, they don’t understand the physics of the real world they just understand the theory of it — but world models do.
World models use words, videos, images, audio and many other sensory inputs to build a simulated reality of the world. This gives AI important context on how the world actually functions through the eyes, ears and body of a real human.
This type of model would theoretically help AI understand ‘human intuition’ much better resulting in an AI that’s proactive and closer to superintelligence.
NATO Deploying ‘AI-powered Cockroach’ Spies
This is 100% real. NATO is working with a startup that equips live cockroaches with mics, cameras and a locally-run AI model (held on a tiny chip) — sending them into enemy territory to map out locations, spy and do recon work.
The cyborg cockroaches excel at navigating across rubble, through tiny cracks in walls and remaining stealthily hidden.
The German military are already deploying these for their missions.
Sci-Fi is fast becoming reality.
Thanks for joining us for another issue. Now go listen to our podcast :)