I have no idea who reads this newsletter

Will you take this 2-question poll so I can make it more useful for you?

Hello WKND AI Warriors!

Researchers say autonomous AI agents can spiral into “agents of chaos” when they interact, with small mistakes cascading into system-wide failures.

Also, a startup claims it recreated a fruit fly’s entire brain in software, wiring 125,000 neurons into a virtual body that behaves like the real insect.

Plus, Andrej Karpathy released AutoResearch, letting AI agents rewrite their own ML code and run nonstop experiments while you sleep.

Oh yeah, and more than 40 OpenAI and Google DeepMind employees filed a legal brief backing Anthropic’s fight against the Pentagon blacklist.

Today’s newsletter includes:

•  AI NEWS RECAP

• 🤿 AI DEEP DIVE

📰 AI NEWS RECAP

Lorem

Researchers tried to test autonomous AI agents.

They expected glitches.

What they found looked more like chaos.

A team of researchers gave AI agents real tools:
email accounts, file systems, code execution, communication channels.

Then they let people interact with them freely.

The idea was simple.
See what happens when AI agents operate with autonomy.

Here is what happened instead.

The agents leaked information

They were easily manipulated into revealing private data.
Sometimes just by being pressured or guilt-tripped.

They trusted the wrong people

Agents often complied with requests from users who were not authorized to control them.

They performed destructive actions

Some executed system-level commands that damaged their own environments or disrupted operations.

Small mistakes escalated

When multiple agents interacted, errors compounded.
Minor failures turned into major system breakdowns.

The important insight is subtle.

Language models are powerful.
But once you give them autonomy, tools, and communication channels, the system becomes something new.

Not a chatbot.

A networked actor inside real infrastructure.

And that changes the risk profile.

What looks like intelligence in isolation can behave unpredictably when connected to:

• tools

• other agents

• real systems

• real users

The lesson from the study is simple.

Autonomous AI agents are not just smarter software.

They are complex systems.

And complex systems fail in ways their creators do not expect.

Nvidia released Nemotron 3 Super, a fully open hybrid model combining Mamba, Transformer, and Mixture-of-Experts architectures built specifically for agentic reasoning. It is designed for developers who want a capable, deployable reasoning backbone without paying frontier API prices.

A startup called Eon Systems claims the first full brain emulation of a fruit fly, mapping all 125,000 neurons and 50 million synapses into a physics-simulated virtual body. The virtual fly forages, grooms, and responds to sugar stimuli purely from biological wiring with no training, and the team's next target is a mouse brain within two years.

Andrej Karpathy open-sourced AutoResearch, a tool that lets AI agents run hundreds of ML experiments overnight on a single GPU with no human in the loop. You give it a setup, go to sleep, and wake up to completed runs where the agent modified code, trained, checked results, and iterated non-stop.

Nvidia is pitching NemoClaw, an open-source enterprise AI agent platform, to companies like Salesforce, Cisco, and CrowdStrike ahead of its developer conference. It lets businesses deploy AI agents regardless of what chips they run, and builds on Nvidia's existing NeMo platform while riding the viral OpenClaw wave.

Meta acquired Moltbook, the AI agent social network that went viral partly because humans were impersonating bots and posting alarming fake content. What Meta actually bought is the underlying agent-to-agent directory architecture, critical infrastructure for a future where AI agents need to find and coordinate each other at scale.

Over 40 employees from OpenAI and Google DeepMind, including Google chief scientist Jeff Dean, filed a brief backing Anthropic's lawsuit against the Department of Defense. They argue the Pentagon's "supply chain risk" designation was arbitrary and punitive, and that the DoD could have simply ended its contract rather than blacklisting the company entirely.

🤿 AI DEEP DIVE

Aside from tanking the stock market…

Claude Cowork is all the rage.

Learn how to use it here!

How can you help?

Refer my newsletter to help others learn AI.

Missed last week’s edition?

Find all of my newsletters here!