In Today’s Issue:
🪰 Eon Systems successfully uploaded a fruit fly's entire brain into a digital simulation
✂️ Oracle is reportedly slashing up jobs to free up billions in cash for $156 billion AI data center expansion
🔬 Andrej Karpathy released a tiny script that lets AI agents autonomously run, edit, and improve their own machine learning research
⚠️ Geopolitical conflict threatens the AI industry's $300 billion desert data center bet as drone strikes hit AWS facilities in the Gulf
✨ And more AI goodness…
Dear Readers,
A fruit fly just made history, not by evolving, but by being digitally resurrected. Eon Systems copied an entire biological brain, neuron by neuron, into a simulated body that moves and behaves like the real thing with 91% accuracy, and that's our featured story today.
But the living world is catching up to the chaos: Oracle is preparing to slash up to 30,000 jobs to bankroll its $156 billion AI data center empire, OpenAI's robotics lead walked out over the Pentagon deal as ChatGPT uninstalls spiked 295%, and drone strikes on AWS facilities in the Gulf just shattered the illusion that $300 billion in desert data centers are safe from war.
Meanwhile, Andrej Karpathy dropped a tiny repo that lets AI agents autonomously run and improve their own ML research: no human coder required. We've also got an exclusive interview with LTX CEO Zeev Farbman, fresh Qwen 3.5 benchmarks that reveal where open-source models actually start being useful, and enough signal in the noise to keep you ahead of the curve. Let's dive in.
All the best,
🤖 Agents Iterating Autonomous AI Research
A new minimal repo called autoresearch by Andrej Karpathy packages the core of nanochat LLM training into a ~630-line, single-GPU script designed for autonomous experimentation. Humans tweak the research prompt (.md) while an AI agent continuously edits the training code (.py), running 5-minute training loops and committing improvements that reduce validation loss. The idea is to engineer agents that can indefinitely accelerate ML research progress by autonomously exploring architectures, optimizers, and hyperparameters while logging results through Git.
🚨 Oracle Plans Massive AI Restructuring
Oracle is reportedly preparing to cut 20,000 to 30,000 jobs to free up $8 to $10 billion in cash flow as it struggles to finance a massive $156 billion AI data center expansion. Several U.S. banks have pulled back from lending, pushing the company to consider selling assets like Cerner and shifting infrastructure costs to clients through strategies like “bring your own chip.” The move highlights how even major tech firms are restructuring aggressively to fund the AI race.
🚨 OpenAI Exec Quits Over Pentagon
OpenAI hardware leader Caitlin Kalinowski resigned after the company’s new Pentagon agreement, citing concerns about potential domestic surveillance and autonomous weapons without clear governance guardrails. She stressed the decision was “about principle,” not people, and warned that such high-stakes AI deals require stronger oversight and clearer limits. The controversy is already impacting sentiment: ChatGPT uninstalls reportedly jumped 295% while Anthropic’s Claude surged to #1 in the U.S. App Store, signaling growing public scrutiny around AI’s role in defense.
Superintelligence exclusive Interview with Zeev Farbman, CEO and Co-Founder of LTX
First Brain Uploaded Successfully
The Takeaway
👉 Eon Systems demonstrated the first whole-brain emulation driving a physics-simulated body - a fruit fly's full 125,000-neuron connectome producing realistic behaviors with 91% accuracy using just four basic components.
👉 This is fundamentally different from AI approaches like reinforcement learning: the digital fly runs on actual biological wiring data mapped neuron-by-neuron from electron microscopy, not trained algorithms mimicking behavior.
👉 Eon's roadmap scales from fly to mouse brain (70 million neurons) and eventually human, positioning whole-brain emulation as a potential alternative path to superintelligence alongside large language models.
👉 The ethical dimension is already live - Eon says it treats the uploaded fly as a potentially conscious entity and is building rich environments rather than test boxes, setting early precedents for digital mind welfare.
A fruit fly just became the first animal in history to have its brain digitally resurrected inside a virtual body, and it actually works. San Francisco-based startup Eon Systems took the complete connectome of a fruit fly brain (125,000 neurons and 50 million synaptic connections) and wired it into a physics-simulated body using MuJoCo. The result is a digital fly that moves, responds to stimuli, and behaves like a real one with 91% accuracy. This isn't a reinforcement learning trick or a fancy animation. It's a literal copy of a biological brain, mapped neuron-by-neuron from electron microscopy data, driving a simulated body through naturalistic behaviors.
What makes this so remarkable is the simplicity. The team only needed four ingredients: the connection graph, synapse-based weights, a map of excitatory and inhibitory neurons, and a basic leaky-integrate-and-fire model. No hand-tuning. No artificial training. Just biology, running in silicon. Previous projects like OpenWorm attempted embodiment with far smaller nervous systems. DeepMind's MuJoCo fly relied on reinforcement learning, not actual brain data. Nobody had closed the full loop before, from connectome to body to behavior.
Eon's next target is the mouse brain, with its 70 million neurons. And eventually? Humans.
Founder Michael Andregg frames this as an alternative path to superintelligence, one where digital minds aren't built from scratch but copied from biology.
Why it matters: This is the first proof that a complete biological brain can be digitally copied and embodied in a way that produces real behavior - turning mind uploading from science fiction into experimental science. The path from fly to mouse to human just got a lot more concrete.
Sources:
🔗 https://www.nature.com/articles/s41586-024-07763-9
Qwen3.5 family comparison on shared benchmarks
Qwen seems to be really useful from 27b upwards; anything smaller than that is noticeably less effective.
War Tests AI's $300 Billion Desert Bet
The AI industry's most ambitious infrastructure bet just collided with geopolitical reality. The Iran conflict is threatening more than $300 billion in Gulf AI spending on data centers, chips, and related investments - and the consequences could ripple far beyond the Middle East. Drone strikes hit multiple Amazon Web Services facilities in the UAE and Bahrain, knocking banking apps offline and shattering the Gulf's pitch as a safe haven for the world's data. Data centers, once seen as neutral commercial assets, are now being treated as strategic military targets - a paradigm shift nobody in Silicon Valley planned for.
The stakes are enormous. The UAE is building a 10-square-mile data center campus, Saudi Arabia plans 6.6 GW of compute capacity by 2034, and companies like OpenAI, xAI, Microsoft, and Google have all placed massive regional bets. Both the Strait of Hormuz and Red Sea - the only two data routes connecting Gulf infrastructure to the world - are now effectively closed to commercial traffic. Think of it as building the world's most powerful supercomputer, then cutting its internet cable.
Gulf states are collectively planning 8–10 GW of AI compute capacity spread across multiple sites, grids, and operators, fundamentally changing the risk calculus. This crisis could actually accelerate the push toward more resilient, geographically distributed AI infrastructure.
