In short: explore versus exploit. Both are necessary, both are valuable.
I largely agree with the OP, because I can only see an explosion in both camps.
The explore side might be automated by compute, and that breaks the social contract of the existing academic incentive structure.
But people who are inherently curious will continue to be inherently curious. Nothing will stop anyone from exploring their curiosity, and the speed and depth at which they explore will only increase. We will build tools that enhance and automate our pedagogical compression.
And for the people who want to cook to feed the family, that automation is coming too.
Or perhaps we'll find a new ceiling: the frontier failure modes that still require a human mathematician in the loop, because discovering some class of novel insight doesn't scale with current AI architecture.
He doesn't make a claim that implies "don't train" might not matter in the way you might think.
They cannot rule out that the data was trained because they have no per-user provenance tracking through the training pipeline once data is de-identified..
The entire point of de-identification is the inability to know the source of data. If the researcher forgot to hit "do not train" then that's that..
The only thing they have is a coincidence and the fact that the LLM may have used the training data that then researcher technically may have agreed to share.
Whether or not that's smoking gun of anything is hard to say. And the fact may remain that the proofs are significantly different, we do not know.
It gets an incredible amount of mileage. And I believe it is the interface that allows the model to learn and eventually bake that intelligence into the model.
Look at how Astra scored 100% on Arc-AGI-3. It was largely because of the harness.
The harness increases the chances (often to 100% chance) of a non-deterministic LLM to perform deterministic actions.
Not only that, the harness provides the feedback that becomes training data for the model. So, over time the model bakes those lessons in, and the harness becomes less necessary and the agent becomes more efficient at some tasks.
A harness will likely always be necessary, we may hit some level of complexity or some level of compute that never allows us to bake the lessons into the model, and external tools provide the model with the ability to find leverage and make up for those short comings.
I did not imply harnesses were useless, just that they're merely a functional part of running an LLM. But it's ticking a checkbox. If the harness can support LLM capabilities (tool calling, reading the screen, visual interaction), it's reached the limit of its potential. Like a JSON parser - can it do the job or not?
Astra didn't solve Arc-AGI-3 because of its phenomenal harness, just a harness was needed to support its capabilities.
Giving Astra's harness to a different LLM would not have improved said LLMs ability to solve the task in a meaningful way.
I feel like this is important to point out, seeing how I'm commenting under an article about how a harness is valued at $50B.
Imagine valuing a JSON parser at $50B on the basis that most communication can't be read without one.
A harness is fundamentally required - although we don’t really see it using tools like Claude or codex ai models still only can take tokens in and produce tokens out. Tool calls are literally just the ai model printing “I want to call search_web with parameters abc” and the harness sees that in the text stream and runs it
"Hey... we need to improve our porn detection AI... but we want to avoid paying for all that porn... is anyone open to taking on some personal liability for a nice bonus this year"
Sell to who? The company is loss-making. Investors WANT the founder to have shares so that the founder is invested in being the force behind making the company NOT loss-making.
My team solved this by creating a PR draft skill that clamps the length of the description to 3-5 sentences max. Those 3-5 sentences must only say WHAT is changing and WHY.
I find it to be far more useful than when humans wrote PR descriptions. Many engineers didn't write one, and those that did were poorly written... this problem is mostly solved for us.. it still has LLMism speak.. but it's useful enough for me to get the context I need to do my review.
The challenge of hard rules like this is that they're always overly restrictive. I've made multi-thousand file PRs that needed two lines of description including the title, and 5 line PRs that needed a 1hr presentation to fully explain them.
My personal guideline is that writing for humans should be done by humans.
I mean.. human judgment still applies. But as an automated first pass the clamp works in 90% of cases. I scan for correctness and make small edits here and there.
I don't follow, the words make perfect sense together in most software engineering contexts.
"Seam" is an industry standard term coined by Michael Feathers in Working Effectively with Legacy Code.
To call a seam load bearing means it's performing critical work for the dependent class, perhaps a database query.
A seam that is not load-bearing would be something that is just injected for testability - maybe a date provider that provides some constant time to avoid flaky tests.
Tbh, this is quite literally the opposite of vapid. A whole book was written about them and their importance, and how to leverage them.
In my experience, Claude uses the word accurately. Code has a lot of seams, and seams are an important thing to communicate when working with code. Therefore, expect to see the word often.
Personally, I don't mind it at all. I'm glad the industry is finally standardizing our language more. Makes it easier for me to communicate with other engineers.
Well, first, we don't have enough context to judge what the "load-bearing seam" was used for in this case ("The key structural point first: the only load-bearing seam is [...]"), so we don't know if it was meant in the well-defined legacy-code sense. But that doesn't really matter.
I'm taking issue with the combination "load-bearing seam". It's a bad metaphor, because seams are usually structural weak points in the physical world, and not load-bearing in the sense that this modifier is usually used. (Seams need to bear loads and stresses to do their job, but so do walls; yet, we do not call all walls load-bearing. We mean something extra when we say that, something that seams don't do.) Even if we were talking about seams in the well-defined software sense, as opposed to the metaphorical one, you still get a mixed metaphor as a result that I find extremely awkward and grating. It doesn't have to be. There are so many ways to highlight the importance of something without calling it "load-bearing".
I understand that you don't see it that way or don't care, but to me, the result is thoughtless, careless and vapid. Bad metaphors put little holes into a text, they leave eddies of confusion where meaning should be, they look load-bearing while actually being weakening, they're like a fart in the elevator that should lift the reader's understanding.
Note that I'm not calling into question that seams may be well-defined in some software contexts, or that "seam" and "load-bearing" can be valid metaphors on their own, as you describe. I think you might have misunderstood me that way. I'm only calling out, and fed up with, the bad style that permeates LLM-generated prose like the whiff of something not quite digested.
It's not this particular case that irks me, but what it exemplifies. I wouldn't mind so much if similar things to this weren't there everywhere, every single day.
In short: explore versus exploit. Both are necessary, both are valuable.
I largely agree with the OP, because I can only see an explosion in both camps.
The explore side might be automated by compute, and that breaks the social contract of the existing academic incentive structure.
But people who are inherently curious will continue to be inherently curious. Nothing will stop anyone from exploring their curiosity, and the speed and depth at which they explore will only increase. We will build tools that enhance and automate our pedagogical compression.
And for the people who want to cook to feed the family, that automation is coming too.
Or perhaps we'll find a new ceiling: the frontier failure modes that still require a human mathematician in the loop, because discovering some class of novel insight doesn't scale with current AI architecture.
reply