iiinova — where your digital self keeps evolving

In fact, precisely the same arguments that used to be used about latent semantic analysis in the 1990s. Although the "latent" there might have been a little more honest.

140

Exactly. A lot of it is just straight knowledge. What Google was completely missing when their AI told me an osprey was a wading bird with a long thin beak.

170

There’s usually a little group of them around the lake where I hike during summer. Beautiful things.

000

This isn’t a diss on that article by the way. Not entirely. It’s more a lament. What will we lose by automating science? By focusing on the surface, on measures and metrics. Is what we lose valuable? To some, probably not. But to me, it was the craft of building science that made it all worthwhile.

020

In short, for me at least, it’s not research at all. It’s a simulacrum of research. It’s Wile E. Coyote research: trebling the effort while the point becomes increasingly distant. Not only is it not my thing, I genuinely have no idea how I’d have mentored researchers to handle it.

I read this and I am glad I’m not an academic any more. It feels like the entire nature of academia, a community, curiosity, inquiry, all of that has been replaced by a drive for productivity, strength, impact. Quality is no longer usefulness, it is a constantly shifting set of metrics.

130

I did look at this a little, because it crossed wires with my work on the ascription of mentality. In the end, I came to believe that perceived similarity is one factor that promotes that ascription. As so often, frameworks like these are fascinating insights into human psychology.

000

I found a great article a while ago which did a deep dive into the actual transcripts from three cases — not quite as catastrophic, but still ending very badly. There were consistent patterns. What this tech needs is some old fashioned qualitative research, but that’s mostly been cut, globally.

030

I’m not using that as an argument for blame legally. As a society, if something like that is a factor in causing harms, we do have a duty to address them and mitigate them. In effect, we need to be able to regulate it, as we do other media.

130

I don’t think reporting is the only issue. There is accumulating evidence that chatbots can, under some circumstances, reinforce harmful thought patterns. That has definitely happened in other cases. So reporting aside, it’s not unlikely the tech is a contributing factor.

Stuart Watt @morungos.bsky.social · 21/02/2026

This.👇👏 How can we have a sense of achievement, of fulfillment, without working for it? We cannot self-actualize for free.

053

I am not sure I can make it there myself, and there are plenty of others who would benefit more and give more than I can, but damn, this is so tempting.

020

Stuart Watt @morungos.bsky.social · 21/02/2026

This looks like an *AMAZING* event! First, COGS and Sussex does outstanding cog sci work. Second, Andy Clark's ideas have blown my mind, positively, on many occasions. I rate his work extremely highly. Third, workshops are awesome to develop good research communities. (And I hate conferences).

111

Stuart Watt @morungos.bsky.social · 21/02/2026

I’d add absence of leadership and strategy, just reacting to events. From my experiences of being the first rat off sinking ships of employment, that was a surprisingly big factor in toxicity.

190

Stuart Watt @morungos.bsky.social · 20/02/2026

If they genuinely think people are going to let AIs have access to their credit cards, they're more delusional than I thought. A few rich bros, sure, but everyday folks on a budget? No chance.

Stuart Watt @morungos.bsky.social · 20/02/2026

That's a possibility, but the only way growth in retail can happen in aggregate is that smaller stores are driven out -- effectively all retail run by a cartel of global megacorporations. It can happen, arguably is happening, I don't see how AI enables it beyond what the internet has already done.

Stuart Watt @morungos.bsky.social · 20/02/2026

Maybe I am naive, but “AI-fueled growth” puzzles me. Generally automation doesn’t do growth — especially when you are already hyperscale. It may cut costs, usually by transferring them, e.g., to consumers. But… growth? Someone will have to explain that to me. It sounds all hopey-wishy.

150

Reposted by Stuart Watt

Timnit Gebru @timnitgebru.blacksky.app · 19/02/2026

Dude who wrote about how "the left is missing out on AI" is on here. Do you see who they are funded by? The biggest EA funders, the longtermist institutes we've been writing about and documenting, including FLI where Muskrat is still an advisor.

Supporters
Our work is supported by a variety of foundations, charities, and individuals who share our commitment to high-quality journalism about AI (grouped by lifetime giving):

$1M+
Coefficient Giving (formerly Open Philanthropy) (2023, 2024, 2025)
Survival and Flourishing Fund (2024, 2025)
$100k – $1M
The Casey & Family Foundation (2025)
EA Infrastructure Fund (2023)
Future of Life Institute (2024)
$10k – $100k
ACX Grants (2024)
AI Safety Tactical Opportunities Fund (2024)
Cullen O'Keefe (2025)
Hazel Browne (2024)
Newman Family Charitable Fund (2025)
Robert and Virginia Shiller Foundation (2023, 2024)

We have also received donations of less than $10,000 from a variety of generous individual donors.

Our donors have no editorial control over the work of Tarbell, our fellows, or our grantees. Tarbell does not accept anonymous donations greater than $10,000. For details, see the Donor relations section of our ethics and standards policies.

16518

Our ability to make inferences about the behaviour of a system is more a property of *us* than of the system. So it is at ourselves and our reactions we need to look.

000

I don’t think that’s it. “What we think of as computing” was itself framed on an abstracted version of human behaviour. There are many computing-like things we’ve had which were very different: cellular automata, GAs, etc. Computing isn’t some magical rational god-phenomenon.

100

The basic setup is that it is driven by ranked constraints that evaluate candidates. Each person can have different constraints and rankings, but they all have to “work” well enough to be useful.

000

So so many, the problem. But if I were you I’d take a look at optimality theory, which intriguingly arose from attempts to bridge connectionism and universal grammars. Much of it is on phonology, sadly, which is quite technical, but it’s also valid for grammar, and it’s quite intriguing.

First, thank you for this service. Second, this thread is gold. It’s a real insight into how HRM’s management is so dysfunctional. But I’m not sure how the city can save itself from it.

050

I mean, that’s slightly flippant, but what we do know is, we do not communicate by predicting logical words to put into sentences. That’s essentially a behaviourist retrospective reconstruction. We now know better.

100

Tom Phillips @tom-phillips.com · 19/02/2026

That’s like saying “how does human language work?” Language is big. You just won't believe how vastly, hugely, mind-bogglingly big it is. I mean, you may think it's a long way down the road to the chemist's, but that's just peanuts to language. (With apologies to Douglas Adams.)

100

Reposted by Stuart Watt

PUT IT BACK. PEBBLE IS NOT FRIEND.

728572

The discussion was not about next-token prediction, it was about human cognition in general, and whether it could (usefully) be reframed via next token prediction. Reading is a reasonable cognitive phenomenon to show that, no, it can't, not usefully anyway.

100

If you have a substantive point to make, then make it.

100

Anything I say would be a distraction from Olivia's analysis of that engineering-oriented framing, so I'll stay out of it. Except, more than a few of us hated the operational interpretations, saw it more as thought experiment about society than science, and saw gender as the key, not a distraction.

050

Oh yes. I was part of a very active scientific community on the TT for a long time, back 1990s-2000s. That said, much of that literature was more nuanced than today's. As a critique of an "engineering-oriented" TT, it's fine, but that was not the dominant framing back then -- in that community.

140

We’ve always used metaphors from tech to explore how people think. The delicious one was the computer, because we ended up with folks using an abstracted version of what (mostly women) mathematicians did as a model of human behaviour. It’s all circular. LLMs as the metaphor is maybe the worst yet.

1324

This, by the way, is how text emerges with bullet points, dashes, bigger spaces between sentences, and using layout. This is not rocket science, it is solid principles from the psychology of perception.

1334

Sal Sawler @salsawler.bsky.social · 18/02/2026

Specifically on this point of next-token prediction, as someone who’s studied this with eye-tracking, people are *not* next-token readers. They skim text. Eyes saccade. Even 1970s AI folks knew this, and old reading systems like DeJong’s FRUMP explored much more plausible mechanisms of reading.

48522

Reposted by Stuart Watt

Ooh, while I’m at it, though, I should also tell you all that @nimbuspub.bsky.social is publishing a 10th Anniversary Edition of 100 Things You Don’t Know About Nova Scotia in April! And just to show how long ago it’s really been since my very first book was published, here’s an old interview!

globalnews.ca

100 things you don’t know about Nova Scotia | Watch News Videos Online

Watch 100 things you don’t know about Nova Scotia Video Online, on GlobalNews.ca

4246

It is not a coincidence that a lot of the markers used by other LLMs to “detect” AI generation are, in fact, rhetorical.

020

I’ll add that LLMs are specifically good at rhetorical forms. I’ve tested this a little experimentally. It is not just bullshit, it is intentionally persuasive bullshit.

It’s as pure clickbait as I’ve seen in a long time. They’re baiting us. We do not need to react.

It’s literally called “Transformer News”, so I doubt if honesty is a priority for them.

Katya King @katyavdgk.bsky.social · 18/02/2026

The interaction with outsourcing will be interesting, because that had already had an impact on those folks. I’ve quite often picked up work rescuing low-budget outsourced tech that turned into an unmaintainable disaster. Vibe coding as personal prototyping I can grasp, but it won’t end that way.

010

Reposted by Stuart Watt

Today's winner of the internet 🏆🏆🏆

011

Reposted by Stuart Watt

THE FAR SIDE DAILY @bot.cowtools.org · 18/02/2026

In its more horrific method of retribution, the mob will sometimes dress victims as mimes, place them in glass boxes, and let them perish slowly in full view of the public.

841790

Yes, definitely. We, together, can be stronger and better than any one group. And the alternative is awful, because it will end up legitimizing oppression. You might not get fascism from the start, but you'll probably make it there eventually.

010

American exceptionalism was the fast track to American superiority, and here we are. It's that global hubris that will boomerang back and destroy them. And in a sense, that is precisely what 9/11 was.

010

As soon as you concede that any one side is intrinsically better, aka superior, you are on the slippery slope to systemic oppression. It is that humility that keeps your state from being an asshole. (A lot of this applies to anti-trans issues too.)

210

I think that's fair. A lot of the risk society literature argues that was a pivotal moment for a whole cluster of reasons. There's a similar dimension to it, too: forcing people to pick only one side. Civic nationalism allows multiple overlapping identities, it has to.

My sense is, and I could be very wrong, that a lot of the abandonment stems from Israel-Gaza, which a kind of Kobayashi Maru of leftist politics. Civic nationalism requires you to welcome and include the oppressed, and they didn’t. Supporting any oppressor denies the rights of the oppressed.