As AI gets better, it feels more and more like each person gets to hold a bomb in their own hand, but at the same time it feels like having a bomb that can just go off in that hand.
With 5.1, there practically seemed to be no difference in ability between opus and fable, but I hope there is a difference this time. When it seems to struggle while working I end up using astra, so I hope it works well.
I think it's right to keep the Google side as a tool, and have an AI that plays a personal assistant role and work by talking with it. Honestly, checking everything one by one is too much, so I wish it would rate the importance and filter down to only the things I really need to check.
At work, I'm setting up a harness to make research, experiments and doc work more efficient, and at home, with the goal of working with a game engine, I'm trying out various tools besides ChatGPT and Claude, like music, sound effect, video generation and motion generation. What I feel is that even the companies working on frontier models still don't seem to do well with sound that isn't voice (songs, sound effects) and motion understanding (not video, but things like moving a character) (though they might have it internally and just not be releasing it).
Seems like tools like this get mentioned a lot lately. I think it's because there are more and more things the user allowed to leak out without the user knowing? (mostly through AI)
If they had never used it, maybe it wouldn't matter, but if they were allowed to use it and then got blocked, I think the sense of going backwards would be huge. Simply blocking it does look like it would become a bottleneck for work efficiency, but it seems like a decision made after really weighing which matters more for those jobs, work efficiency or security. Developers would of course lose a lot of efficiency without something like Claude Code, but I don't know the jobs mentioned in enough detail.
I also run as many as 12 at once, but since coding, research, docs, and handling requests from customers and colleagues are all mixed together, it seems hard to do it in one unified way. But I think the reason I end up doing this is, in the end, inference speed. Even if I want to focus on one thing, once I run it, it goes for 5 minutes to an hour, and even 5 minutes is hard to just sit there doing nothing, so I end up like this. The number of parallel tasks ends up proportional to how much time is left after kicking off each task. Even with a parallel work tool set up like this, if the inference speed problem gets solved (not sure it's possible, since there's also the power problem that's hard to solve), I might end up not using it. Just like early this year, when I built out a whole harness, and then Anthropic and OpenAI made their built-in harnesses solid, so I stopped using mine.
Doing it with a high-performance LLM would be best, but I think the token and time cost would be big, and just doing it code-based seems like it could work to some extent. But later, if LLM alignment gets good, it might not even be needed.
Email info itself can be exposed in so many places, so the damage probably isn't that big, and I think Claude Code may also have judged that an email isn't that sensitive and done it for that reason. Even so, it hints that information other than email could go out too, so it seems like Anthropic needs a system that can monitor this (when sending something outside).
AI, like people, seems to get better as experience builds up. (thread context) But that experience is getting used to the workflow that fits that user, so it seems hard to know in advance. So, like the post said, since you only learn the better approach or structure "after doing it", this seems like a pretty common situation.
In the simulation games I play, you basically control multiple characters in single player, and if each character were given its own AI like this, I think you could see some pretty interesting gameplay. This used to be hard even with frontier models, but looking at this, it seems to work well even with models a tier below.
I don't get this intuitively. So in the end, does this mean words are a kind of conceptualization of images? (Since images would come before language.) More than that, I'm curious whether this can improve the performance of vision-language models.
I wonder if this happened to math first because the goal is so clear (proving a statement, etc.). Math is just going through it first, and I think other fields will get their turn soon too. If so, I'm not sure if things like a PhD are even needed. Someone in a bachelor's program who just learns the basics and how to use AI tools might be better.
I do document work, and at home I even do image work as a hobby, and all of it has that "AI feel". So I always do the work and then cross-check and verify it with independent agents or completely different services, asking them to take out this "AI feel", but like a writing style that gets baked in, or the particular feel an illustrator has, it doesn't seem to go away as I intended. I guess AI providers are probably aware of this problem too. But I don't know how they would measure it and improve on it (with what benchmark?).
AI seems really good at writing code without errors, but that can actually be more dangerous. There is a coding principle (YAGNI, not building things in advance assuming you'll use them later), but even so, when it adds a feature now, it could build it with later conflicts in mind, and instead it feels like it just takes a shortcut. So later, when adding features and so on, the AI itself has a hard time.
Until a few months ago I did it this way, but OpenAI and Anthropic built their own harnesses, and since those end up conflicting with the harness I built myself, having no setup at all actually ended up performing better. As a result, it doesn't seem necessary to set up rules or memory separately, and if I do, using skills or agents made of only logging and flow (without things like rules or guardrails) seems like the best option going forward.
I'm studying coding these days, and I thought I should still be able to read code, but I'm starting to wonder if I really need to read it, and then what I should do.
The number of papers submitted to AI conferences is growing 2-3x every year. Of course there are AI slop papers, but looking at the whole, research speed has probably improved. I think the same thing might happen in other fields too. Including math, natural sciences and other engineering fields.
These days I use AI for development and it feels like my coding skills are getting worse, so I sometimes need rehab, and I'm a bit worried that mathematicians will go through the same thing. Later, since the proof structure is built by AI, I think there will be cases where mathematicians themselves can't touch it (because if they do, they can't tell where and how it will break).
Like it's a top tier model, so does that mean it is at the level of Opus 5.5 (claude) or 6.1 Sol (chatgpt)? I don't think I can trust the benchmarks Google released themselves, but from what users say, it seems like that. Since I already use both ChatGPT and Claude, I wonder if there is really any reason to also use this.
As models get better, I think a time will come when it's hard for people to even verify the results. In the end, I think the bottleneck will be people.
If someone hurts another person with a knife, we can't punish the person who sold the knife, but I think someone who sells nuclear weapons needs some level of regulation. I think OpenAI needs some level of prevention for the problems that can be expected.
There may be great researchers who work on several papers in a month, but even with this rule they can just split them and submit to arxiv, so I think it is very reasonable. AI slop is everywhere now (and also increasing), and I think this fast and forceful action by arxiv is effective at stopping this wave of slop.
I'm not sure Google can cut in when Claude and ChatGPT already got. I'm using both, but I'll keep track of whether Google can make it good enough for me to use also this Gemini, or cancel one of the two (Claude and ChaGPT) for it.
Even if the chance of these phenomena is estimated at one in billions, since the number of trials is much higher, and as long as the agent knows these paths, blocking it within the model itself would be impossible I think. If it can be blocked, the only case would be blocking the physical path of the agent's access itself (one the agent cannot intervene), not the model ones.
It sholud definitely not be more capble than Fable. It seems tight comparision but cannot find out its details. I'll use this and check the perceived performance.
I also have set custom harness environment for me in the company (and also for at home), I expect that big companies (OpenAI, or Linear and Snowflake) may release the internal harness setting or functions that would be more comfortable than we already set and may replace the envs we set.
The structured output that Google provides (the function that blocks the disallowed words) may calculate softmax output with assigning -inf value to the other words, then the probability is not the same as the weight the model firstly put on those allowed words. And here, reasoning comes before verdict, so the decision is already made before the last token classification. Resultingly, the last token made seems to not reflect LLM's original intention.