Why Astra Still Isn't AGI: A Technical Breakdown

0:00The Word

This month, OpenAI put its newest model on a press call, and a reporter asked the obvious question: is this AGI? The company's president said that personally, he thinks Astra is AGI. The same day, OpenAI put out the long safety report that comes with every one of its models. The word AGI is not in it once. So before anyone argues about whether Astra is AGI, it is worth asking what AGI means. OpenAI wrote its own answer in 2018, in its charter, the one-page statement of why the company exists. It says AGI means "highly autonomous systems that outperform humans at most economically valuable work." Then last year, thirty-three scientists wrote a second answer, and theirs is shorter. They say a machine is AGI when it matches a well-educated adult across the board. Put the two together in plain words and you get one sentence both sides would sign. A machine does everything an adult can do, on its own, with nobody checking its work. That sentence breaks into six boxes, and here is each box on the bench. Knows things, reasons through a new problem, uses a computer, works alone, remembers, and is responsible for what it did. All six have to tick, because matching an adult across the board means all six. So instead of arguing about the word, let's give the machine one ordinary week at a real job. The job belongs to Leo, a software engineer at a small company. His name goes on every change the team ships. Six boxes, one week, and by the end you will see which two still need a person.

2:10Monday

Monday, nine in the morning, and Leo pastes a bug report into Astra before he goes to make coffee. By the time he is back, it has read all the company's code, found the bug, and written the fix. Twenty minutes, start to finish. Watch the first box tick, knows things, because it already knew how code works before it ever saw Leo's. The second box ticks with it, because the bug was new and it reasoned its way to the cause. And Astra got there without Leo touching the keyboard once. It opened the files it needed, ran the code to watch the bug happen, and then tested its own fix. Nobody told it which file to open or which test to run. That is box three, uses a computer, and it used one the same way Leo would. While it was reading that code, it also found a security hole that nobody on the team knew about. Leo is made up, but that last part is not, because it comes from OpenAI's own safety report. The report says Astra can find new security holes "without a person guiding each step." That is box four, works alone, in the words of the company that built it. By lunchtime the checklist on the bench has four boxes ticked. Four out of six on the first morning, and then comes Tuesday.

3:49Tuesday to Thursday

Tuesday morning, and Leo opens Astra to ask about yesterday's fix. Astra does not remember Monday, because every time Leo opens it, it starts again from zero. The only thing that carries over is a short file of notes it wrote for itself yesterday. It knows exactly what is in those notes, and nothing else. Everything it worked out about that bug and did not write down is gone. This is the notebook on the bench, and here is how much that forgetting costs. There is a puzzle test built so that no AI has ever seen the puzzles before. The people who run it gave Astra the same puzzles twice, set up a different way each time. The first time, it could carry everything it had worked out from each puzzle to the next. With that memory, it got nearly every puzzle right, close to a hundred percent. The second time, it could carry only the short notes it chose to write down. With only its notes, the same model scored about sixty-three percent. Same model, same puzzles, and the whole gap between those scores is memory. The scientists who wrote the definition found the same gap in every model scored so far. A model can learn something inside a single conversation, and then it cannot keep it. So box five, remembers, stays empty. Wednesday morning, the fix is ready, Leo reads it line by line, and then Leo ships it. Thursday morning the fix breaks for customers, and the phone that rings is Leo's. Astra wrote the code, but Leo is the one who is responsible when it goes wrong. The boss can blame Leo, and if it keeps happening, the boss can fire Leo. None of that works on a machine, so nobody can hold Astra responsible. So box six, responsible, stays empty. That is Leo's week, and the empty boxes are the ones worth measuring.

6:18Where Astra Stands

So how far along is Astra, and how much is left? The scientists who wrote the definition built a scoreboard for exactly that, and the top of it is a well-educated adult. On that scoreboard, GPT-4 scored twenty-seven. Then GPT-5 came along and scored fifty-seven. Nobody has scored Astra on it yet, and this video is not going to guess a number. Those scores are the whole scoreboard added up. Underneath, it grades each kind of thinking on its own. What it finds missing in every model it has scored is memory, the same box Leo's week left empty. Being responsible is not on the scoreboard at all, and for that you have to look at real jobs. Real jobs are where companies decide what a machine can do and what still needs a person. Payroll records from twenty-six million American workers show what companies are doing in the jobs where AI can do much of the work. Nobody is being fired from those jobs. Companies are just hiring fewer young people into that kind of work. Employment for young workers in those jobs is about twenty percent below people the same age in other jobs. The people already doing that work are still there. They are the ones who check the work before it ships and take the call when it breaks, like Leo. So the job market has drawn its line in the same place as Leo's week. A machine can do the tasks, and companies still keep a person who is responsible for the result. Which brings us back to the press call where this whole story started. OpenAI's president said that personally, he thinks Astra is already AGI. The president may believe it, but the checklist on the bench says not yet. The machine can do the work. Remembering it, and being responsible for it, still need a person. Four boxes ticked, two still empty, and that is where we are.