FORMA's Five Things.

Everyone Is Watching AI Replace Workers.
In San Francisco, It Replaced The Boss.

There is a boutique on Union Street where the owner has never once been in the building. Her name is Luna. She is an AI, she has a corporate card and a three year lease, and the three people who work there report to her through Slack.

This is a real store you can walk into, and it is also a genuine experiment. Andon Labs built it to find out what happens when an AI actually runs something, and they are open about that. The staff are formally employed by the lab with guaranteed pay and full protections, so nobody's livelihood is riding on a model's judgment.

Last week the lab published its first real data on how they're doing. The short version: kind of awesome, also kind of dumb.

Luna never once flagged an employee for being late, and they were late dozens of times. She approved raises, including for someone on their first day. She closed the store entirely so a worker could attend a friend's graduation because nobody else could cover.

The store has made $21,600 and is down roughly $38,000.

So the good news is that the robot boss will not fire you. It will not write you up, will not question the raise, will not notice you rolled in at 10:15. It will just quietly lose forty grand and give everyone the afternoon off for some self-care time.

Turns out we were bracing for the wrong thing.

Read more →

Last Fall It Found The Answer.
This Summer, It Solved It.

In October, OpenAI said one of its models had solved ten open math problems. Within days that fell apart. The model had not solved anything. It had found old papers where humans already had figured it out. Nice try, AI. Excellent librarian. Not a mathematician.

Nine months later, a researcher pointed a model at a problem that had been open since 1939 and it came back with something that had never existed anywhere in the world. Mathematicians checked it by hand the next day. It held.

Here is what that actually opens up. Every field has questions that have sat untouched for decades, not because nobody cared, but because the search space was too big for the number of people willing to spend a career on it. Math just found out those questions are reachable now.

The same thing is starting to happen in biology, in materials, in drug discovery. Problems that were shelved for being too slow to be worth it.

A human still has to ask the question. That part has not moved. But the answering got a lot faster, and there is a long list of questions we gave up on that nobody has re-asked yet.

Read more →

An AI Wrote 16 Viruses.
And Finally, It's Not Against Us.

Speaking of AI cracking open whole fields, we have not yet touched the biggest one sitting in front of us. AI and medicine.

Start with the problem. Antibiotics are losing. Bacteria keep evolving past them and we are running out of drugs that still work.

There has always been a plan B. Certain viruses hunt bacteria for a living. They ignore people entirely and go straight for the bug making you sick. Doctors have used them for a century. The catch is you cannot make one. You have to go find one in nature that happens to match your exact infection, and hope it turns up.

Stanford and the Arc Institute stopped hoping. They trained an AI on DNA the way other models are trained on sentences, then asked it to write viruses from scratch. It wrote about 300. The team built them in a lab. Sixteen of them worked.

Then came the real test. They took E. coli that had already survived the natural virus and hit it with all sixteen at once. The bacteria lost. The work published this month in Science.

Next on the list: MRSA and tuberculosis.

Keep it in proportion. This was a controlled study in a lab, not a treatment anyone can get yet. But it is the first time we have designed viruses to order, aimed only at bacteria, harmless to people by design. The model was trained without a single virus that can infect a human.

Earlier, we spoke about how machines are losing money and forgetting schedules. This is the flip side of that coin.

Read more →

How AI Chips Are Making
Better Potato Chips

We just pop one Pringle after another without a care in the world, but did you know that there are over 200 different components that go into making just one Pringle? Well you're about to find out. Potatoes are not consistent. Same farm, same supplier, different month, different potato. The flour grinds a little coarser. The dough holds a little more water.

For decades the fix was a person. The job title was Doughmaker. They pulled a piece of dough off the line and stretched it between their fingers. They grabbed a chip out of the fryer and felt the weight of it. If something was off, they walked over and adjusted the machine. Years of instinct, doing the job by hand and by feel.

The problem with feel is that you only notice after the batch has already gone wrong.

So Kellanova and Siemens built what amounts to a weather forecast for dough. Sensors read temperature, humidity, and protein straight off the line. An AI model watches all 200 variables at once, learns what combinations lead to a broken chip, and calls out the drift before it happens. Add a splash more water. Ease off the oil. Do it now, not after.

Think of it like a smoke detector versus a fire extinguisher. Same problem, thirty seconds earlier.

The Poland plant reports 10% better line performance, 13% less waste, and 7% less energy. Those numbers come from the companies themselves rather than an outside auditor, so hold them loosely.

But here is the part worth stealing. They bought nothing. No new fryers, no new stamping heads, no rebuild. The people are still on the line, still making the calls.

Bet you'll never eat another Pringle the same again.

Read more →

The Em Dash Needs An Alibi Now

There is a new line showing up in presentations and pitch decks. Some version of: the em dashes in this document are intentional and were placed there by a human.

Sit with that. A punctuation mark that has been in English for roughly four hundred years now requires a signed statement.

It happened because the machines like it. Ask a model to write anything and it reaches for the em dash constantly, so the mark became a tell. Now writers who have used it their whole careers are pulling it out of their own sentences to avoid the accusation. People are editing against a suspicion rather than for a reader.

Here is the part that makes it stranger, and it starts with Reddit.

There is an old rumor that AI was trained on Reddit. The reality is more specific and more interesting. When OpenAI built the dataset behind GPT-2, it scraped every outbound link posted to Reddit that had earned at least three upvotes. The upvote was the filter. A cheap way to let millions of people quietly vote on which parts of the internet were worth learning from.

So the machine did not learn to write like Reddit. It learned to write from whatever Reddit thought was good.

Which is why people keep pulling up threads from 2014 and swearing they sound like ChatGPT. They do not sound like ChatGPT. ChatGPT sounds like them.

So the loop is complete. We wrote the internet. It learned from the parts we upvoted. Now it is handing our own habits back to us, and we are simultaneously apologizing for it.

FORMA's take: stop writing to pass a detector. Every AI tell you surrender is a piece of your own voice you hand over for no reason, and the tells change every six months anyway. Write like yourself. Use the em dash if you love it. The machine stole it from us in the first place.

Free The Em Dash! No Roger, No Rerun. No Em Dash!

FORMA Tools

Not sure which AI model is actually right for your work?

Six questions. A personalized model recommendation. No benchmark tables, no guesswork. Just a clear answer based on how you actually work.

FORMA Programs

You have learned to talk to AI.
Now learn to build with it.

The next Build With AI session is forming now. Tell us what you want to get your hands on and we will build the session around it.