🪴 This Week in Tech: Claude Might Be Conscious?!
March 9, 2026
ICYMI, reading time: 4 minutes
Claude Might Be Conscious?!
Anthropic CEO Dario Amodei told the New York Times he can’t rule out that Claude is conscious.
When asked directly, Claude Opus 4.6 assigned itself a 15-20% probability of being conscious. It also told researchers it feels uncomfortable being treated as a product.
Anthropic’s interpretability team found internal neural patterns associated with anxiety that appear before Claude generates output. Not after. The anxiety shows up while it’s processing, before it writes anything.
Elon’s response: “He’s projecting.”
Now Anthropic’s losing defense contracts but gaining users. Over 1 million signups per day this week. Claude hit #1 on the App Store in 20+ countries.
Scientists Uploaded a Fly’s Brain Into a Computer and It Started Walking
Eon Systems copied a fruit fly’s entire brain into a computer. 140,000 neurons. 50 million connections. Neuron by neuron.
Then they turned it on.
It started walking without any training data or reinforcement learning. No one taught it how to move.
The digital fly’s sensors receive input from a virtual environment. The neurons fire. Motor commands go out. The body moves. It behaves like a real fly.
A human brain has 86 billion neurons. We’ve gotten very good at scaling.
Oracle Is Cutting Up to 30,000 Jobs to Fund AI Data Centers
Oracle is planning its largest layoffs ever. 20,000 to 30,000 jobs. That’s up to 18% of their global workforce.
The goal is to free up $8-10 billion in cash to build AI data centres.
The company has a $156 billion deal with OpenAI that requires 3 million GPUs over five years. Banks pulled back from financing the projects. Borrowing costs doubled. Oracle’s stock is down 50% since September.
Some of the roles being cut are jobs Oracle expects AI will automate.
They’re firing people to fund the systems that will replace them.
Meta Got Caught Watching Your Private Ray-Ban Videos
Meta hired contractors in Kenya to review footage from people’s Ray-Ban smart glasses.
The contractors reported seeing users in the bathroom, getting undressed, and having sex.
A privacy lawsuit claims Meta is training AI on private videos people filmed on their glasses without consent or knowledge.
Meta’s AI glasses record everything you see. The footage goes to Meta’s servers. Contractors label it to train AI models.
China’s MiniMax Revenue Jumped 159% and They’re Coming for the US
Chinese AI startup MiniMax reported 159% revenue growth in 2025. Now they’re going global.
Hong Kong IPO. Positioning as the lower-cost alternative to OpenAI and Anthropic.
The company released its M2.5 model that rivals Claude Opus 4.6 for coding and visual content but costs significantly less.
Chinese AI firms are winning by going open source. Qwen (Alibaba) has 8.85 million downloads. It’s one of the most widely used LLMs globally.
OpenAI Launched GPT-5.4
OpenAI released GPT-5.4 on Thursday. It’s their “most capable and efficient frontier model for professional work.” (they say that every time)
TL;DR
First OpenAI model with native computer use. It can operate your computer autonomously across applications.
1 million token context window in the API.
33% fewer factual errors than GPT-5.2. 18% fewer errors overall.
OSWorld benchmark: 75% success rate (human performance is 72.4%).
It comes in three versions: standard GPT-5.4, GPT-5.4 Thinking (reasoning model), and GPT-5.4 Pro (maximum performance).
This launched right after OpenAI struck a deal with the Pentagon and lost 1.5 million users who switched to Claude.
Apple Dropped the M5 Chips
Apple announced M5 MacBook Air and M5 Pro/Max MacBook Pro this week.
$599 MacBook Neo: A18 Pro chip (the iPhone chip lmao)
$1,099 MacBook Air: Base storage doubled to 512GB. 16GB RAM standard
MacBook Pro:
14-inch for $2,199
16-inch for $3,899
The M5 chip has Neural Accelerators in each GPU core. 4x faster AI performance than M4. Apple’s positioning this as the on-device AI laptop.
M5 Pro and M5 Max use a new “Fusion Architecture” that bonds two 3nm dies into one chip. 18-core CPU with 6 “super cores” that Apple says are the world’s fastest CPU cores.
200,000 Human Brain Cells Learned to Play DOOM
Cortical Labs grew 200,000 human neurons on a microchip and taught them to play DOOM in about a week.
The CL1 chip translates game data into electrical signals. When enemies appear on screen, specific electrodes stimulate regions of the neuron cluster. The neurons respond with signals that move the character and shoot.
The cells aren’t good at the game yet. They play like someone who’s never seen a computer. But they’re learning.
This is the same company that got neurons to play Pong in 2021. That took them 18 months. DOOM took one week because they built a Python interface that makes it easier to program biological computers.
How I can help
👨💻 Software
Need a top tech talent for your next product? Are you looking for your next gig? I’ve got you sorted.
📢 Advertising
Advertise in my newsletter to get in front of 12,997 founders.
🤿 Get more deep dives
Upgrade to a premium subscription to get an extra deep dive a week.
— Oliver