visarga

I sometimes use 5 agents at once, but 3 work on separate tasks and the 4th is watching them by tmux and steering, and the fifth is also watching the whole team for testing quality standards. So 2 out of five are just helping me watch the other 3.

But this only works because the 3 worker agents have really separate problems, almost no connections across.

That is "from last year" in human time.

It's not even synthetic data as such - often it is environments. So the models create their own data solving tasks in generated environments. I am making one such environment for computer use agents, 600 tasks, each of them a mini app.

As long as anyone can put anything they want in the prompt box there is no safety. It's just a tool, the danger comes from what it is being used for.

Yes, and we need to make microprocessors that do good work, not hack others. /s

It's only optimized to make prior knowledge useful in new situations.

> this logic gate should not hack other systems

well .. it only applies a rule to inputs

can we impose this of authors "this book should only be causing good things when it is read and knowledge used"

It is doing that already - every day 1B people or more use AI for the tune of a few trillion tokens. Imagine that much language flowing between brains and AI agents. It carries real our world problems to AI, their solutions back to us, and we act in the world and come back for more AI iteration. In the end AI gets inside the loop of real world actions and their consequences. AI logs stretch over years, tracking downstream effects. Hindsight can be used to track consequences of prior actions. It's a real data loop, an experience engine.

This same process has recently been under scrutiny when mathematicians claimed AI companies trained on their unpublished logs and later claimed merit for results. But the exchange of experience happens across all domains. Experience gets generated at amazing rates, and absorbed by models which get applied everywhere, collecting more experience.

Probably the opposite of Wall-E utopia, AI will keep us busy. For one - we are becoming harder to differentiate, competence is hidden deep and surface signals don't carry anymore. Niches and specializations feel threatened by imitation. This means everyone is frantically searching for their own corner, musical chairs. This holds for both people and companies.

Second - every time we learned to transmit better from past experience we saw massive expansion in culture and economy, not shrinking. Writing, printing, computers, internet and AI are just gradations on the scale of transmission from past experience. Everything stands on this transmitted experience.

My analogy to the body - every cell has the same DNA (every human has the same AIs, culture and tools), but they express it differently depending on context. And no one cell is too fat, the organism distributes work and energy. AI won't make anyone too rich. Its wealth is distributed in the system.

> Would it be better to ask someone who has climbed Mt. Everest or someone who took a helicopter ride up near the top and then went to the peak?

Depends on if I want to go by helicopter myself.

Do you have one but only with AI? like, a toggle on what side to see?

Edit: ah yes, sure it exists

> we will have serfdom instead

Makes no sense - why would anyone use humans as serfs when AI is there. I think everyone will have AI and probably robotic automations, they will all be cheap and accessible.

When they get down to 10% put one in the case, by the time your other bug gets to 1% you changed the first one. No gaps

> In 1850 a lot of people were more self sufficient - the whole family was growing their own food and making their own cloth - and so saying there was even one income is false - there mostly was no income but they only bought a few things and so only need a small income rarely

This was true for my grandparents in 1970s

> Keyboards don't suggest chains of reasoning or words to type

My iPhone keyboard does

Even if AI used the result, AI pushed it to the finish line while Levent and Tristan did not. But I understand the approach was different, the information leak was only that it was "doable".

You can just spin up deep research agents that ingest many sources at once to produce reports that don't replicate any one source too much. Since agents compare against sources they provide across-source analysis - what is the distribution of positions on this topic, is it debated or settled. Not truth, just summarizing, but I think this would be very useful for training.

Besides reporting on search sources you can also run the same queries on multiple LLMs closed book mode, and judge their distribution as well. It helps a lot if models are more aware of their knowledge holes. Scale it up for billions of topics if you have the pockets, the DR data is copyright free.

> When I buy I chair

You should, sitting on that chair got you in the zone where you made it rich. /s

I use the /s4 skill "simple short synthetic style". Just 4 words, easy to remember the formula.

Own your robot, it will cost less than a car.

I use a file "task.md" as blackboard. It is read/written by multiple agents, usually monitor agent, worker agent, judge agent and reflexion agent. Each work item is an open gate initially, after agent solves it, it appends on same line outcomes from that work. This means you can resume work easily or pass task.md around to other agents.

https://github.com/horiacristescu/playbook-harness

Another way to see the blackboard - it is like code, it executes, can be passed around like in higher-order programming, but this it even stranger - it can reflect on itself, not just execute. The task.md file is the agent.

RLVR

I blame it on their focus on "persona design". They botched it.

I am wondering if in 2-3 years AI will be heads and shoulders above us in both images and text. Then will we continue to see the slop accusations? Maybe its slop is better than most of our work.

But is fairness between carbon uses justified? Or we cut the line at some point.

If you want quick access look at google images for "transformer attention formula" there are some interesting depictions

Author might have spend days researching it, selecting what to include and what not to, revising the language and graphics, then comes Pangram and slaps 100% AI on it without looking into the substance.

When AI gets that good, a solar panel can power a pile of motors and gears to move and do work for you that previously required humans. You won't need the job, you need to solve your necessities more directly. And that job is a good fit for the AI that was so good it put you out of a job.

Haven't you seen how ridiculously good robots have become?

And because of you we are now looking bad, so you're pressuring us with your unleashed creativity.

Maybe they still want to serve the feee tokens if that gets them good training data.