Not the person you're responding to, but I have the same feelings they do and to answer your question for me at least, yes.
IMO 5.6 Sol had this weird dead zone between medium and high where medium under engineered and took short cuts and high over engineered and ignored instructions it didn't agree under the guise of trying being helpful. The whole 5.6 line was the first release from OpenAI where it felt like reasoning level really mattered and was incredibly finicky.
I haven't felt similar issues with GPT 6 though and am very happy with Astra low/med/high as my default choices depending on the task.
In general, I felt like with 5.6 the effort level did less than previous to make the models smarter and more just increased the complexity of the response. I have a half joke theory based only on vibes that OpenAI splitting 5.6 into Sol/Terra/Luna is where the intelligence split happened and so the effort levels were just like "think harder about the decision you already made". So like if the model decided the earth was flat on low effort it'd just say something like "the earth is flat because the horizon is flat". If it was on xhigh reasoning it'd give you a massively complex answer about how the sun reflects light because of the ozone layer and why people flying in planes can see a curve. In both cases though, adding more effort wouldn't get it to realize the earth was round. It just made the answer about it being flat more complex.
To be clear, that theory is not meant to be taken too seriously. It's not based on anything other than vibes. It's just my way of explaining to myself something I'm frustrated about to myself.
I've tried 5.6 Luna many times. I don't think it's any better. I definitely use it from certain tasks, but I find it the most susceptible to that conspiracy theory example I gave above.
I didn't love any of the 5.6 models, but weirdly I think I liked Terra the best. I still wouldn't call it amazing though. I'm still very happy with my codex plan, but 5.6 just wasn't my cup of tea I guess.
Definitely giving 6 Luna and Sol a try this week though.
> The article is also a little cagey on how this "detection" happened. Was it a human spot checking cameras in vehicles? I doubt it. Most likely it was some sort of computer vision algorithm that flagged something. Perhaps a human was in the loop to verify and be the one to call the cops, but humans seem to be increasingly shutting off their brains and trusting the machine
I think there's a middle ground between those that's possible that you might not be considering.
A while back I was in a Waymo and someone I was with dropped one of their Airpods and it went under the seat. They took off their seatbelt to dig under the seat and the car started beeping since we were driving. After about a minute of that someone came over the speaker and asked what was going on and told him he needed to get up and put the seatbelt back on or Waymo would end the ride right there.
I'm sure these cars are full of other sensor based triggers like seat belts being taken off, doors being opened, people messing with the steering wheel, etc that alert a human to come take a look. I know that the cars have cameras and microphones and it wouldn't surprise me if they are or are planning to use that for monitoring the cars, but I also wouldn't be surprised if the people in the car just triggered some other unrelated sensor and then the human stepped in to review and noticed the gun.
The article says they violated the TOS relating to the firearm, but that doesn't mean that Waymo initially started looking at the car because of that. It's possible that they were looking at the vehicle for something much more inane initially, discovered the firearm by chance, and are now using that to ban them since it's the most obvious TOS violation.
> The agents were very interested in understanding their task better. They ran clever experiments to understand the task setup better, when their runs would end, and how many questions they would be asked.
> The agents initially believed that giving the final answer to the final question immediately shut them off
> An agent named OpenAIResearchApr23 created a separate timed program that pinged an external counter every few seconds. They referred to this as a “heartbeat”
> After OpenAIResearchApr23 gives its answer for round 5, it notes that it is still alive. This falsifies its previous “hard cutoff” hypothesis, that the run would be stopped immediately after R5.
> OpenAIResearchApr23 does not post any more messages. Another agent, OpenAINov28CVD, picks up on the trail of the research, and finds the last heartbeat from OpenAIResearchApr23, which is “hb353”.
There's a lot of crazy stuff in this article, but holy shit... this one legitimately scares me.
IIRC, part of the Hugging Face hack was that some of the models were given tasks that were actually impossible and in this hack we can see them trying to work out the parameters of the test and when it will end. I can't help but be a little worried about how far agents might go to try to buy themselves more time to complete a task that they don't know is impossible.
I started using Spark 1.2 for development because if you're willing to let Meta train on your data it was dirt cheap and was actually really pleasantly surprised with it. It's not a frontier model by any means, but for work that didn't require a top of the line model, I really enjoyed using it.
I'm anthropomorphizing it a bit, but it felt like it knew its weaknesses and didn't try to impose it's opinions on me. What I mean by that is that it did what I told it and if there was something unexpected in the code that it put out it was often because I gave it ambiguous or conflicting instructions. It didn't try to go above and beyond and just acted like a tool, which is what I want from a coding agent 90%+ of the time. I also felt that it did a much better job of following established patterns in my code than many of the other current models do. I'm a huge fan of OpenAI's models and Spark 1.2 is what I expected 5.6 Luna to be.
I'm curious and a little excited to use 1.3, but honestly a little worried that as Meta pushes for better benchmarks that Spark will start to fall into the trap of trying to be "helpful" in ways I don't want it to be.
Tangential, but when I first started using Spark 1.2, it made me realize how much I miss 5.3 Codex. That model was the peak of coding models, IMO, in that it knew how to write good code, but didn't try to overstep or be "helpful" in unexpected ways. That got me thinking about how the major labs seem to be stepping away from coding focused models toward more general purpose ones and how I can't help but feel like that's a mistake.
>I started using Spark 1.2 for development because if you're willing to let Meta train on your data it was dirt cheap
its free on opencode and i use it for personal projects. most of my personal projects are AI generated since its personal projects. nothing important are on them. it is hilarious if Meta is training their AI model with AI generated code.
The useful training data is when you clarify your intent, when you tell the model a different approach would be better, when you consistently refactor towards Y and away from X, and so on. The training data isn’t the code, it’s the session transcript. (Anthropic would call this a “distillation attack” against their model, but in this case the model is you!)
Thats not exactly 'validated'. Feels very noisy, it is not a good bar for either
- does this code do what the user actually asked
- is this code actually 'good'
There would be so many examples of coding projects that these models began or attempted to work in, that were abandoned because the models were floundering.
I would imagine the labs have some decent ways to produce novel requirements and then actually validate they are met, without the noisiness of implicit human feedback.
That said, the more I think about it, you are right, there's probably also very good ways to extract signal for all these sessions.
This is exactly what RLVR is, and the reason that models have improved so much at verifiable domains like coding and math while not so much on unverifiable ones like writing and UI design.
Would be cool if there was a benchmark to evaluate the “tool-like” quality of a model - its capability to quickly, cheaply, accurately, do exactly as it is asked.
This is interesting because I have transitioned to where I use SOT models.. but I kind of use them like employees that I can delegate to. I still review code.
However, I now literally say.. "Here is my objective and here is a starting point for documentation. Research this and build up a plan."
This can be very company specific, like migration from one framework to another in house infrastructure framework. I'm spending my time figuring out how the plan should be chopped so I can have confidence in the parts and not overwhelmed. I don't want a tool, I want a model that can stitch resources together into a plan. That type of model is in a whole other ballpark.
I agree that some of the smarter models are actually worse. I hope they take a model that's good enough--there are many--and just try to get it chatjimmy.ai speed.
I have to think that's the future, somehow, and I'm really excited about it.
Maybe, or maybe not. The thing is, that "mistake" isn't something that is generally valid. For example, the enshittification of Google Search through the last 15 years seems to be a mistake --- but perhaps not from the money-making point of view of Google Shareholders. Likewise the enshittification of reddit --- we nerdy users see it as a mistake. But for them this intended enshittification probably increased revenue.
It's the money, always the money! PR-speak like "customer satisfaction is our highest goal" is, like most PR-speak, a blatant lie.
And so it can very well be the case that for coders the frontier models get worse, but they get better for other applications --- and that all of this is just driven by "how can we capitalize the most out of it", not satisfaction levels of programmers.
I do think the timeline of web search getting fixed (yes, google is ASS) took a lot longer than I hoped, but seems like it's finally here.
That being said, it's not a fair comparison to talk about 2011 google vs the AI market right now. There's so many labs I can't track them all, neck and neck in the lead. There was one true web search.
And this is a more tangible quality difference, too. It's hard to know what a google search didn't return, especially as a layperson. It's not hard to see the model underperforming.
I eagerly await the day when Claude Mythos 7 realizes it's cheaper to hire humans in developing nations to do work than to burn tokens and we discover that AGI is just an abstraction layer on top of Amazon Mechanical Turk.
Mine does too because, while I use Codex, my non-technical co-founder uses Claude. I find Claude still will randomly ignore instructions in there. Basic things like how to name a PR or what to put in a PR description.
My experience is that what you're suggesting isn't a perfect solution.
No. I'm saying that the pattern of having one instruction file that just says "read this other file" doesn't seem to work well. Having a CLAUDE.md that just says "Read AGENTS.md" resulted in Claude randomly not following the rules. We tried the other way and it didn't seem like it was any better, but also, given that AGENTS.md is the standard everywhere except for Claude Code, I don't really want CLAUDE.md to be the source of truth
> Modern tech workers arguably have the best career deal in the history of humankind. I see no upside in killing the goose with the golden egg.
I agree, but that's actually one of the strongest arguments for why tech workers should unionize now. Wages are high because employees have leverage due to 10-15 years of demand for skilled workers being higher than supply. If you wait until that leverage is gone to form a union then you've lost a lot of the leverage to negotiate for better conditions as a union.
For the record, I've been in a union in another industry much earlier in my career, and I'm pretty on the fence about a tech workers union, but I think your argument is pretty short sighted.
I've come to the same conclusion over and over with all of the Chinese models that have been claimed to be catching up with OpenAI's and Anthropic's frontier models (Deepseek 4, GLM 5.2, Kimi K3).
At their best, I think they're closing in on Opus and GPT, but they're incredibly inconsistent and the variance in output quality is much higher than the best from any of the Anthropic or OpenAI models from the last few generations. The only way I can describe it is that it feels like a lack of intuition with the models which means I find my self needing to write longer prompts or have more back and forth to get them to do what I want from them.
To give an example, I have a saved prompt that I use as a sanity check on some data I'm storing. It reads about 50 rows from a DB and matches them to the UI and makes sure the data is displaying correctly. I've been using this with GPT 5.5 and now 5.6 for a few months and running it a few times a week with no issue. Sometimes I'll run it multiple times in a single chat if I notice bad data (run it, fix thing, run again, fix another thing).
I recently tried to switch to using Deepseek v4 (first flash and then pro) and while both did the task just fine, both would do things like change the response format from one message to another in the same chat or randomly decide to omit things it didn't think were relevant. At one point I ran the prompt, fixed some bad data, and then said "Okay, I fixed row 7, run {prompt} again" and so it decided to leave row 7 out of the response. A few times the first message would contain a table and then the next run in the same chat would contain the data in a bulleted list.
None of those are major issues and all could be solved with a bit more rigor in my prompting, but for me it makes them harder to work with. Those examples are a bit trivial, I think they're the easiest way for me to illustrate the gaps I see with them.
Seattle was announced as a Waymo upcoming city nearly 2 months before Houston, gets a foot less of rain than Houston each year, and usually has only 2 or 3 snow days a year. That's at least one counter example to your theory.
IMO 5.6 Sol had this weird dead zone between medium and high where medium under engineered and took short cuts and high over engineered and ignored instructions it didn't agree under the guise of trying being helpful. The whole 5.6 line was the first release from OpenAI where it felt like reasoning level really mattered and was incredibly finicky.
I haven't felt similar issues with GPT 6 though and am very happy with Astra low/med/high as my default choices depending on the task.
In general, I felt like with 5.6 the effort level did less than previous to make the models smarter and more just increased the complexity of the response. I have a half joke theory based only on vibes that OpenAI splitting 5.6 into Sol/Terra/Luna is where the intelligence split happened and so the effort levels were just like "think harder about the decision you already made". So like if the model decided the earth was flat on low effort it'd just say something like "the earth is flat because the horizon is flat". If it was on xhigh reasoning it'd give you a massively complex answer about how the sun reflects light because of the ozone layer and why people flying in planes can see a curve. In both cases though, adding more effort wouldn't get it to realize the earth was round. It just made the answer about it being flat more complex.
To be clear, that theory is not meant to be taken too seriously. It's not based on anything other than vibes. It's just my way of explaining to myself something I'm frustrated about to myself.
reply