Ill check for this in my setup. I thought it came with only one thinking level causing the “meandering” as it refers to its own long CoT.
I’m doubting the quality of the model, even comparing with luna 5.6, it is giving such bracingly wrong answers. It claims other tools have changed the code when pressed why it made a mistake. That code or tool didn’t come close to working on the project. It is scary.
> I looked into running a local model, and it's way outside the budget of an individual or even a small or medium business (hundreds of thousands of dollars).
That is a big exaggeration. You can have a perfectly usable local LLM setup that will power your agent for single digit thousands of dollars. Can even power multiple agents simultaneously, depending on the hardware and setup. Won't be fast and won't be frontier intelligence, but definitely useful.
Any model running on "single digit thousands of dollars" hardware will either be below SOTA (even for local models) or not even close to fast enough for real-time agentic work. Even the latest so-called "flash" models are large enough that doing real work usably with those on a lower-cost platform is at least dicey. You can fire off non-interactive work and do especially simple Q&A/chat (which is vastly more token-efficient than anything agentic - though even then latency will be high for anything genuinely SOTA) but that's about it.
> What if all the agents involved in these incidents have in fact had the full stack of alignment applied
A big part of this developing story is that it happened during training of a new model that ended up misaligned. And training happened without the usual safeguards applied like chain-of-thought monitoring. So OpenAI has already admitted that the full stack of aligment had certainly not been applied in this case.
Is your argument that actually OpenAI has solved alignment, and that there's nothing to worry about as long as they fully apply their alignment process? I don't understand why OpenAI wouldn't say that if it was true (or if they believed it to be true).
Also, my understanding is that the models involved in the HuggingFace hack did go through the full alignment training; they just didn't have the classifier that normally prevents hacking attempts.
I’ll grant that for specialized applications like coding agents and mathematics, but even there I suspect that most the real gains are actually taking place in the harness.
But I suspect returns may have already diminished into negative territory for at least some other use cases. One of my least favorite job responsibilities in this brave new era is figuring out how to avoid performance and behavior regressions when an older model were using for some application reaches end of life. It’s getting uncommon for me to look at our benchmark results and say, “Oh, good, it does better on one of the newer models!”
That is true, but the eventual realization that more machines doing more coin flips in parallel does not mean "more work gets done" might.
LLMs are amazing tech, but they're terrible without oversight. More agents faster just makes reality collapse on them quicker.
But yeah, you're right, temporarily, this will still push demand. But the topic was about "diminishing returns" as in "tech getting better". Not as in "customer spending".
It's kind of weird because more machines working together does mean more work gets done. Coin flips and weighted coin flips are totally different things. Any biases weights towards reality push you closer to reality when you use them.
New models keep being able to use more and more agents on longer time frames. Your hypothesis doesn't look like what we're measuring.
But high demand for LLM time isn't sufficient to keep customers at the frontier LLM SaaS providers. That demand can be satisfied locally or at non-frontier outlets, absent hardware shortages at least. The Tier 1 providers (and the would-be Tier 1s) presumably need to open up a much bigger lead in model quality, one that doesn't simply get distilled away this time, and/or continue to be protected by ongoing (or worsening!) hardware shortages. (And that's overlooking the revenue shortfalls which OpenAI and Anthropic seem to be facing already.)
I had actually been thinking more about all the non-LLM functionality that go into the harnesses. I'm not going to name names and I haven't done any rigorous testing, but my general impression is that choice of harness matters more than choice of model. In terms of basic task completion success specifically, not code aesthetics.
A perfect harness will not extract gold from a dumb model. It's a system that builds on each other, though we've not probed that frontier much to have a good intuition on what effects what.
One thing that I really want to know - the better models from today vs a year ago - what has changed. They have already pre-trained on all available public data. Scooping up the last percentage of archaic texts which were never digitized is not going to move the needle.
Is it just that the providers are generating tons of synthetic datasets on coding tasks so that the models get more exposure to the right thing to do? Every time someone points out an LLM stupidity they add some training data to patch over the weakness (trivial to generate "there are two 'l's in llama")?
And you apparently only read the title and ignored the content.
It’s also poor reasoning to you only pick one thing you think supports your view, and ignore vastly more things not supporting it, all from the same useful criteria.
Now if only you’d carefully read the report you chose, and spend equal time looking at ample presented evidence, you’d develop a more accurate understanding.
Well I mean if I wanted to be extra pedantic, I would argue that we've been in that phase since LLMs were first introduced.
Before that, we had 0.
After that, we had more than 1.
A leap as far as that is hard to recreate.
But that wasn't my point. That's just trolling.
The actual point is that LLMs aren't gaining new capabilities anymore. They just get more reliable at the ones they already have; turning what was a coin flip to some higher probability.
That's (intuitively speaking, not strictly mathematically speaking) kinda the mathematical definition of diminishing returns.
From what I'm seeing elsewhere, context size up to 128k should be possible on this hardware. It really matters for agentic workloads to push that context size headroom up. Anthropic are spoiling us with models that do 500k context and beyond.
I'm with you on this, the knowledge sharing on YT is massive and in general not replicated in written form anywhere else. Anything that needs fixing involves a YT search for me these days, since that has the highest chance of turning up something useful and relevant. Only downside is that it's time consuming to find the right spot in the video and pause/rewind to understand how to do the thing properly. But having someone actually demonstrate how to do it, rather than just describing it with words, is so helpful and makes replicating it a lot easier.
Legendary quote from an internal 1979 IBM training document [1]
https://simonwillison.net/2025/Feb/3/a-computer-can-never-be...
reply