This is Scale & Strategy, the newsletter that catches you up like a blue shell in Mario Cart.
Here’s what we got for you today:
Biological chips might be AI’s weirdest — and most interesting — escape hatch
Your AI sounds off-brand? It’s not the model. It’s your lazy brief.
Biological chips might be AI’s weirdest — and most interesting — escape hatch
Tucked inside a pretty normal office in San Francisco, one startup is trying to solve AI’s energy problem with something straight out of a sci-fi script: living neurons.
The Biological Computing Company (TBC) came out of stealth in February with $25M in seed funding and a very non-Silicon Valley idea. Instead of pushing harder on chips, they’re building compute on top of biology. Real human brain cells, grown and wired into chips.
Running the show are former neurosurgeons Dr. Alex Ksendsovsky and Dr. Jon Pomeraniec. Not your typical “ex-Google PM builds AI startup” situation.
Inside their Mission Bay lab, the setup looks like a crossover between a biotech facility and a hardware shop. The team is small, about 23 people, but stacked. AI engineers, physicists, biologists, the usual Avengers lineup. Each chip holds somewhere between 100,000 and 500,000 living neurons.
And no, this isn’t just for show.
Here’s the play: real-world data like images and video gets encoded into these neurons. The system then decodes that biological activity into richer representations that feed AI models. In Ksendsovsky’s words, “all of the data flows through the biology.”
Under the hood, the chips use multi-electrode arrays to interact with neurons grown from stem cells, reprogrammed into brain-like cells. They used to rely on rat neurons, which sounds like the start of a horror movie, but they’re moving away from that. The chips last about a year and, like anything alive, they create waste that needs regular cleaning. So yes, your “computer” now needs maintenance like a houseplant.
This whole approach is basically a return to AI’s roots. Neural networks were originally inspired by the brain, then drifted into pure math and silicon. That shift made systems powerful but also brutally inefficient. Scaling today is mostly brute force. More GPUs, more energy, more problems.
TBC’s bet is simple: biology already solved efficiency.
Early research backs that up. Models trained on biological neural signals hit peak performance up to 3x faster, with fewer training cycles. Translation: less compute, less energy, same or better results.
For now, they’re not selling the chips themselves. They’re using them to improve AI systems, especially in visual domains like video generation, game rendering, and computer vision. Behind the scenes, they’re already talking to major model labs and security firms.
And they’re not positioning this as some distant science project.
“We’re building products now,” Ksendsovsky said. The goal is to prove this works commercially, not just academically.
The bigger ambition is where things get wild. TBC wants these biological chips sitting directly in the inference loop. Real-time compute powered partly by living cells. Timeline: maybe five to ten years, assuming they solve a few minor issues like automating waste removal and figuring out how to keep cells alive indefinitely.
So, small details.
Look, the whole thing sounds slightly insane. But so did AI not that long ago. Now it’s chewing through power grids and forcing the industry to rethink energy at a global level. When the bottleneck gets that big, people start reaching for solutions that used to sound ridiculous.
Fusion. Space data centers. And now… brain cells.
If TBC is even partially right, the next leap in AI might not come from better chips.
It might come from something a lot closer to home. Literally.
Your AI sounds off-brand? It’s not the model. It’s your lazy brief.
Most people blame the tool when the output feels generic. Reality is simpler and slightly more embarrassing. You gave it “write conversationally” and expected magic.
AI doesn’t read between the lines. It reads the lines. Literally.
If you want it to sound like your brand, you need a brand guide built for AI, not humans pretending to be helpful.
Start with your voice, but actually define it. “Conversational” is useless on its own. Spell it out. Is it “talking to a friend over coffee” or “polished but relaxed operator who knows what they’re doing”? Those are very different outputs.
Then give examples. Not vague guidance. Actual side-by-side comparisons.
Show what good looks like. Show what bad looks like. Show what “almost right but still wrong” looks like. That’s how you close the gap.
The most valuable section ends up being the simplest: do this, not that.
Not just tone, but specifics: Use short sentences vs long-winded ones Prefer concrete language over buzzwords Write like a human, not a quarterly report
And yes, set hard rules. The stuff you never want to see again.
No “leverage.” No fake inspirational lines. No robotic openers that sound like a LinkedIn ghostwriter had a stroke.
AI is weirdly obedient when you’re this explicit. It just needs the rules spelled out like you’re talking to someone who takes everything at face value. Because it does.
If you want to go one level up, write full comparison paragraphs. One on-brand, one too stiff, one too generic. That’s basically cheat codes for consistency.
Then actually use the guide properly.
Upload it. Reference it. Tell the model to follow it. Don’t just attach a doc and hope it absorbs vibes through osmosis.
And don’t treat this as a one-and-done thing. Update it. Add examples when the AI screws up. Save prompts that work. Kill the ones that don’t.
The payoff is fast. Your output stops sounding like everyone else. Which, given how much AI content is flooding the internet, is kind of the whole game now.
