astrobiased

This is an exciting time, but also there are very real changes at hand. Game changing discoveries will likely not happen at universities but at private AI frontier labs. Moreover, the models that are used on this are not always open to the broader community today.

What will it be like in two years? Will such models that help enable paradigm shifts in science be only accessible to employees of those private companies? Still thinking about the ramifications ... amazing and discombobulating at the same time.

Appreciate the input! Responses below to your first two items:

1) I'm leaning more into a broader sense, which is that given the priors a system already possesses, how efficient can it acquire competence on a novel task? If I'm reading the point you make, you're focussing on continual learning right? If so, I'm not necessarily restricting my statement above to that.

Here's another reframing: How much of the benchmark improvements come from overwhelmingly large training distributions vs improving the models for adapting to things genuinely outside of it?

2) Excellent points about crystalized and fluid intelligence. Wouldn't the LLM scaling gains be a representation of crystallized capabilities? In regard to Gf, that is exactly what I am asking about. That is what seems to be lacking, Gf like adaptation under genuine novelty.

My concern is that it's increasingly difficult to tell of what looks like Gf like behavior is really coming from better adaption vs. having broad priors from the model's large learned distributions.

I think the moat that China has is energy costs. It's taking learnings from the Bitter Lesson. If you role up scale and compute to the next level, it's energy resources. China has it and sharing open weight models is an effective means of removing the tech moat. This idea has been floating around for a bit now (I'm not taking credit for it).

I can’t help but notice how much this echoes Francois Chollet’s On the Measure of Intelligence: https://arxiv.org/abs/1911.01547

Most of frontier-model progress still looks like skill acquisition optimization: broader benchmark coverage and performance, more domains absorbed into the training distribution, and increasingly strong performance within that surface area.

It seems more about coverage-driven competence. Somewhat analogous to overfitting at scale.

The harder question, in Chollet’s framing, is: how efficiently can a system learn to do something genuinely new?

With our current AI architectures and training in place, I think we will only continue on skill acquisition optimization vs. truly novel intelligence.

Depends on the type of agentic task though. For simple operations, a small model can be quite beneficial.

Not RL. SFT.

Yes, used bert model with decent results.

Pi does one thing that I love, developing a tool that has minimalism where it's easily configurable with good documentation. The leads to new use cases that the the author(s) would have never dreamed of. The organic growth process of the Pi ecosystem has been fascinating to observe. It's one of the reasons why Pi has become one of my favorite coding agents to this day, flexible beyond personal uses and extensible to larger environments.

IMO, I view it more than a coding agent, it's a coding agent platform with powerful extensibility.

Is this in any way similar to Goodfire's work? https://www.goodfire.ai/research/rlfr#

The post hits it spot on with unequal access to the models in terms of security. I'm developing OSS where security is important for the user ... but the frontier models like GPT 5.6 and Fable flake out and state that I cannot get the info/access.

This is extremely lopsided I'll have to resort to GLM 5.2/K3 to ensure that those security issues (hopefully) are resolved properly.

For OSS, this is one of the most counterintuitive experiences I have ever had. More than ever I'm convinced that open weight and open pipelines models are 100% critical for progress on the AI and societal fronts.

Same, so much in fact that I have used it for my website now and it works beautifully for graphics and formulas.

It's the right direction, but control flow introduces limitations within a system that is quite adaptable to dynamic situations. The more control flow you try to do, the more buggy edge cases that pop up if done poorly.

Still have yet to see a universal treatment that tackles this well.

John is not only smart and knowledgeable, but an incredibly great person to know in general. I worked with him on a project briefly back in 2012 and he stood out as a champion for science, coding, and education. His posts clearly reflect him well.

Absolutely love these type of keyboards. But ... with how much security I work with for logins, etc, the fingerprint button on my Mac keyboards are amazing time savers that I don't want to live without. Has anyone found a workaround?

What I find odd is that definition of "investor" is not that clear. When you click through the links you get blocked at the data provider with no context. There's also a link to another post by the same news provider. When clicking through reference to the data source, the link doesn't work.

It would need to be more than that. A prompt for one model can have different results vs another. Even when the model has different treatment for inference, eg quantization, the same prompt for the unquantized and quantized model could differ.