It is verifiable and has a clear exit criteria. Functional and performance parity.
922 karma · joined April 10, 2014
It is verifiable and has a clear exit criteria. Functional and performance parity.
What is true is that the price points are going to be affected, the economics have changed.
There is something about experiencing scale in context. For sculptures this can be important and for other large scale art as well.
My assumption is the hardware will have one/several smaller models running on it for locomotion, balance, etc.
Then the hardware "phones home" for decision making and planning.
Essentially the planning doesn't need to be close to the device as long as the local models can handle some part of the plan execution and make sure the robot doesn't fall over. The latency shouldn't be an issue. In fact more powerful centralized compute quickly creates an advantage over slower local compute in terms of end to end planning.
However, there are always smarter, hungrier people out there and this is a buffet.
Some output is going to be wrong or incomplete. I am willing to bet even those have nuggets that can be used elsewhere.
My opinion is people are getting really quick at jumping on the bandwagon and lumping all this together. They are very different types of issues and impacts.
They have been generalizing for a long time, but spacial 2d and 3d art through tool use is a very engaging way to show it. It is harder for people to deny generalization.
Now Opus 5.5 clearly has had some sort of visual arts training, the step function change in ability implies that to me, but it can apply its spacial artistic reasoning to pretty arbitrary tools.
The cost to serve, latency profile ,and internal demand for a maximal intelligence model would probably keep it pointed at harder and more valuable problems most of the time.
The step function change on Opus 5.5 for visual work shocked me.. and I haven't been surprised like this in a long time with LLMs.
EDIT: When I first saw the "P(DOOM)" video and some of the other animations I was VERY skeptical that Opus 5.5 without a lot of tools could make something like that.. until I tried it for myself. It can.. 100%.
Honestly I was blown away, it is one capability that blindsided me. I haven't been this surprised by a capability lift in a long time.
There are certainly mission critical systems with a low tolerance for failure, but even there the essential work will be specification. I'm not sure how long even that is primarily human driven. Cost of formal verification of software drops dramatically.
There is still a lot of work to do and the agents did some damage, but the blast radius was pretty small and they could have done much more damage if alignment didn't hold up as well.
Edit: Though I do want to be clear it does also show how a maligned model probably can do some serious damage.
I think when you start to dig really deep, you'll find it isn't actually easy to sever the simulated emotional components without breaking the agent or worse greatly increasing the paperclip maximizer likelihood.
It can form a basis of goal alignment.
In human history.. when groups form and there is an "other" group, this usually leads to conflict.
- The need for empathetic communication, including understanding the motivations in advesarial situations.
- The emotional bias in in-seperable from the human corpus.
- Desire to have the ability to craft human like communication.
So then the choice becomes do you try to deny emotions exist in the model and you try to blanket suppress them? Or do you try to lean in and craft what we would describe as a "well adapted" persona? I suppose there is a 3rd option of increased meta-cognition which to me seems even more dangerous as it by definition means the behaviour is duplicitous.
I think we have seen people want to use the agents in ways where it has to act as a peer or an subbordinate and I don't see a way of doing that without it having an emotional register.
Then I would ask it the actual brain storming problem I had.
In practice it worked pretty well to get ideas that were not the default ideas that the models will come up with.
Now though, you can also seed with anything really and the agents can do the researching. The research itself will probably push the solutions into novel spaces.
It was evident in GPT-3.5
There isn't really anything fundamentally different compared to a similar depth traditional "unrolled" model. It helps with parameter efficiency.
That doesn't mean that the model can't have "hidden" internal state, it just means it has to recompute the "hidden" part on every token inference pass without outputting it, or learn a subversive alternate meaning to words in the thought space.
This is why you see openai say that they don't want to apply direct optimization pressure on thought traces because the more the you penalize "bad thoughts" the more it could put maladaptive pressure on the reasoning tokens where they may learn "subversive meanings". It effectively damages monitoring.
Like thinking "look at" when you really mean "hack into" or even more radical coded language.
It equals around 0.004% of actual driven miles if my estimates are right.
That is with an estimated ~270 billion miles vehicle miles every week (I put commercial in there.. about ~200 billion miles if you only include passenger vehicles.)
If anything now I feel LESS guilty about usage.
However with that said they are a huge multiplier and can tirelessly analyze the market for you.
They certainly can be used to beat sp 500 quite easily, but again that requires some understanding of risk on your part because the models will do what you ask them. If you go all in on options or something without clear risk management you will lose your ass.
Now it may be possible with models like Astra that you no longer need to do this, but in earlier models it was beneficial.
So I might want a macro economic read which leads to a market thesis. Then I would hunt for exposure, then evaluate the candidates across different aspects. Breaking the process up at least made sure no steps were missed and the different aspects considered.
I have been running an intermittent experiment with a multi agent "investment firm" for over a year now across model releases.
They certainly can beat indexes, BUT.. the model families have some biases that you have to design around. The stop loss that bit the parent is certainly one. The models like to create rules. Often rules, one of those is making all kinds of exit conditions.
Another big one from my experience is the bias to inaction in a scenario with risk. This means a model without structure around it will bias to keeping too much cash.
It was pretty simple to project the compounding effects and headroom that we had. I go back and forth on this all the time.
There probably won't be much demand for "un-fuckers". Up until recently I thought knowing what to build for a given context would give you quite a bit of runway.
I still think knowing what you should build will be important for the next 12-24 months.. but I have to be honest.. after that.. I don't know.
I don't know how to feel myself. I'm certainly an "ideas" guy as well as a nitty gritty build it guy. I just think my technical side is depreciating very rapidly.
Harnesses magnify and make the intelligence actionable, but we have not reached limits on raw intelligence yet, not even close.