iiinova — where your digital self keeps evolving

Scott Lindsey @scottlindsey.bsky.social · 14/10/2026

The linked conversation is relevant. If you don't use these tools you will probably miss out on startling jumps in functionality. For instance, you might expect an AI Expert to be aware that frontier models can indeed reason their way through simple math problems now. bsky.app/profile/emil...

010 0

Scott Lindsey @scottlindsey.bsky.social · 13/10/2026

Kind of seems like a “straw-person” argument.

000 0

Scott Lindsey @scottlindsey.bsky.social · 11/10/2026

By this argument very few humans are conscious.

060 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

We don't have to guess as to whether or not what LLMs do is protected fair use. www.forensisgroup.com/resources/ex...

U.S. District Judge William Alsup granted Anthropic summary judgment on its use of the named authors’ books to train the large language models at issue. The court found the training use transformative because the books were used to develop systems capable of generating new text rather than to distribute substitute copies of the authors’ works to readers.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

LLMs do not have a literal memory of the training data. That seems like a foundational misunderstanding.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

When I write code, is that not directly derived from my training data?

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

This is not the case. A model trained on millions of pieces of code does not need to call out each piece of code it was trained with, each time it emits new code, anymore than you would have to cite every piece of code you ever read when you write new code.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Perhaps you should not be forced to accept reduced quality. But the assertion that LLM output is always of lower quality than human output is not proven, and may already be incorrect. The more likely outcome is that by crosschecking with LLMs we will reduce the rate of medical errors.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Seems like the "reversal curse" has been fixed btw. But explain who is be "forced to put faith in an illusion"? Someone who uses a tool that I create with an LLM?

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Sounds like a distinction without a measurable difference.

Giant robots labeled "AI" roam the landscape while humans huddle around a fire, and a human says "But they're not actually 'reasoning'".

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

We measure "understanding" the same way a school teacher does -- by testing to see that the student has generalized concepts and is not simply parroting. Mathpocalyse should be evidence enough: github.com/openai/math/... We know that they think because we can literally look at them doing it.

A thinking trace from an LLM showing the model's "thoughts".

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Cognitive computer? I'm not sure what that means to you. As long as we are dropping new terms, I like "half-mind", which is comes from The Mercy of Gods by James S. A. Corey. I think it captures what AI does pretty well. Thinking without learning, etc.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

No, we're ok with the fact that the robot with pistons achieves walking in a very different way than a human would. Walking is the observed output. We're smart enough to know it won't get tired and won't benefit from an energy drink, etc.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Uh, yeah so no. Open Source is all in on vibe coding. Check out our founding father. en.wikipedia.org/wiki/The_Cat...

210 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

if it was ingesting the water this would be true. Is it ok if we say that a robot can walk?

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

I think that “understanding” is something you measure through observation. If it doesn’t bump into the wall, it understands. Any other definition is probably a back door to believing in souls without saying so.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

For a real deep-dive you could read Masley, but my take is that saying that a well-configured Roomba "understands the room" is perfectly valid. And LLMs, what they do is pretty similar to "understanding" in humans unless you argue that only meat can understand. bsky.app/profile/andy...

210 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

A novel solution. I'll be sure to give credit when I find myself using it.

110 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Well if it's not in the "Chinese Room", then it's making some kind of connections between the symbols it handles... I don't know what word to use if "understanding" is verboten.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Wandering into theory of mind here, far deeper waters than what I can speak to. All I will say is that if hold on to the "chinese room" theory of how these things work you won't find it to be very predictive. Anthropomorphizing them on the other hand works out pretty well.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Yes I can be in favor of AI and agree that gas turbines in neighborhoods are, actually, bad. These things are getting conflated and Elon made it so much worse.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

The resultant artifact is in fact a tiny sliver of all the training data by size; reading my code would probably only very slightly increase the models understanding of syntax. You won't find most small projects "remembered" in the model at all. It is more like reading than inclusion.

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

Yeah no one uses Grok, but it's to such a degree that Elon has taken to renting out his terrible datacenter. I'm generally pro-datacenter but neighborhoods have got to keep an eye on externalities and be willing to say "no".

100 0

Scott Lindsey @scottlindsey.bsky.social · 08/10/2026

As someone who's code was presumably scraped, that code was not stolen. I still have it. Reading open source code (or anything else placed on the internet) is legal for very good reasons.

220 0

Scott Lindsey @scottlindsey.bsky.social · 07/10/2026

You use AI to analyze it. You ask an agent to adversarialy review it and you ask another to review the review. If you really want to understand the details you read it and ask AI to help you when you get stuck. This is the essential skill of the AI age.

A still from The Ballad of Buster Scruggs in which two people are being hanged, one is labeled "mathematicians" and the other "coders" The person labeled "coders" says "first time?"

100 0

Scott Lindsey @scottlindsey.bsky.social · 07/10/2026

The general process is to paste a fragment of a work you think a model may be overturned on, and see if it continues it. However, It only works if the content was duplicated in the training set, as was Harry Potter (thanks to kids seeding it around the internet). arxiv.org/abs/2601.02671

arxiv.org

Extracting books from production language models

Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized dat...

110 0

Scott Lindsey @scottlindsey.bsky.social · 06/10/2026

And always they think the law is a slam dunk in their favor, but instead of retaining a lawyer they are posting their feelings.

020 0

Scott Lindsey @scottlindsey.bsky.social · 06/10/2026

But I am not a lawyer. If you think you have a case you should talk to one?

010 0

Scott Lindsey @scottlindsey.bsky.social · 06/10/2026

The headline is fairly misleading — this is a narrow ruling that doesn’t change the general rule that it is legal to train on copyrighted material. The rules that allow derivative materials such as study guides remain intact. www.supersummary.com/old-mans-war...

supersummary.com

130 0

Scott Lindsey @scottlindsey.bsky.social · 06/10/2026

I mean, Microsoft did try to license the right to train, even though there was no legal requirement for them to do so. Blame Disney for starving the public domain if you like. bsky.app/profile/scal...

000 0

Scott Lindsey @scottlindsey.bsky.social · 01/10/2026

Yeah I called Opus out after it finished a “one day” task in 18. I think it just has no frame of reference for its own speed.

000 0

Scott Lindsey @scottlindsey.bsky.social · 29/09/2026

Agreed, to all of this. I think academics are just not comfortable with a "student" learning things that they aren't specifically told. They wants the AI to know exactly what they tells it, and superior performance from an emergent system is just a horror as well and a threat to their prestige.

110 0

Scott Lindsey @scottlindsey.bsky.social · 29/09/2026

Right? I think it's just the culture of symbolic AI, all that work they did quantifying rules into Lisp, decades of effort for very little return. When LLMs showed up with the implicit message that all the effort was wasted... Just too much to process. chatgpt.com/share/6abb5e...

111 0

Scott Lindsey @scottlindsey.bsky.social · 29/09/2026

> I'm baffled by these PR stunts It will start to make sense when you accept that "stochastic parrot" is a terrible frame of reference. Understand, LLMs majorly upset the AI priesthood -- the idea that meaning can arise spontaneously from language is career breaking for the former elites.

130 0

Scott Lindsey @scottlindsey.bsky.social · 27/09/2026

Famously Joe Pantoliano did not like rare beef so they had to use shiitake mushrooms for some of the shots.

040 0

Scott Lindsey @scottlindsey.bsky.social · 09/09/2026

The allegation is that conversations between the leads and Codex may have made it into the training data and given the agent some ideas about approaching the problem. But actual mathematicians think this is real. Check out Terrence Tao's essay over on Mastodon if you want.

030 0

Scott Lindsey @scottlindsey.bsky.social · 08/09/2026

The fallback position has been identified btw -- the next argument will be that the results were stolen from Buckmaster and Alpöge, conveniently not mentioning that the two proofs are unrelated and solve the problem in very different ways.

151 0

Scott Lindsey @scottlindsey.bsky.social · 08/09/2026

If I read her correctly, she just doesn't believe it is possible, hence "extraordinary claims require extraordinary evidence". Solving Millennium Prize Problems is not a thing that a "stochastic parrot" should be able to do and she's not able to grapple with that.

270 0

Scott Lindsey @scottlindsey.bsky.social · 17/05/2026

If you really wanted to get a deep understanding of a project, a coding agent would help you get there faster. However, you’re right, typically you’re not gonna bother. Developers will be expected to support more projects thanks to better tooling.

130 0

Scott Lindsey @scottlindsey.bsky.social · 17/05/2026

Personally I have used Gemini CLI and OpenAI Codex for this sort of work. Claude Code is probably the original for this task but that was just one more subscription than I wanted to pay for.

020 0

Scott Lindsey @scottlindsey.bsky.social · 17/05/2026

From there, you want to get your linting and typechecking in order. These really help agents stay on the straight and narrow. With all that in line, you want to start making improvements. Look for duplicate code, architecture errors, etc. Basically everything I would do normally, but faster.

110 0

Scott Lindsey @scottlindsey.bsky.social · 17/05/2026

First off would be booting up an agent with the codebase and just asking questions, lots of questions. The first task it to try and get it running on local, if that is how it is meant to work. Once I had a baseline, the task would be to add as much testing as possible with coverage reports.

110 0