iiinova — where your digital self keeps evolving

I didn't say all human research is insightful or creative! But at least human researchers have a point of view, some sense of direction, and are taking some risks when they try to publish their work

000

The final output is something that may appear to have insight, creativity, or whatever, but if it actually had those things *it would not have needed* to generate thousands of other dead-end proofs and run tons of pointless simulations, etc

100

Unpack what actually happens: 1) It's prompted, often with expert direction 2) It can search to pull in context about the specific problem, so transition probabilities are conditional on a lot of information provided by human experts 3) It burns through thousands of dollars until it gets lucky

100

And if the transition probabilities have been learned from ~all text on the internet and ~all books, it's not surprising (and hence certainly not impossible) for the random symbol sequences to respect the same geographic constraints that exist in the world that generated the text

000

I don't see what you're disagreeing about. Randomly stitching together words from a biology textbook has some probability of generating hypotheses that turn out to be scientifically useful.

200

These are good for "verifiable domain" problems. And there will be some scientific/economic value found by generating orders of magnitude more code (though I doubt it will be anywhere near enough to justify current valuations) I don't think Stochastic Parrot critiques ever said this couldn't happen

100

Some benchmark results imply that the likelihood of (5) has increased. Maybe that's good, but it doesn't mean we're no longer dealing with Stochastic Parrots. The same software product that can code (extremely inefficiently) can also generate more persuasive-sounding astrological charts

111

Possibilities: 1) It hasn't actually done any ("independent") checks 2) It did check, but the check was wrong 3) It did check, correctly, but the report of the results is hallucinated 4) Check and report correct, but there are still other errors remaining it didn't find 5) All correct

110

Claude now tells me something like: "Three independent check rounds each found errors in my [Claude's] drafts" OK, inference compute go brrr. How much should I trust a Stochastic Parrot that "checked" its own work... stochastically... and is--again, stochastically--reporting the results?

110

This enabled large improvements on many benchmarks. Much wow, very frontier Does this mean that the SP criticism has become irrelevant? I don't think so. Here's one example to explain why:

120

I would say, in increasing importance: (1) LLMs are bigger (~10x parameters), (2) "harnesses" and "tools" automate some of the context curation problem, (3) "inference time" computation -> generating many responses and automatically picking the best

210

Re: the Stochastic Parrot (SP) arguments It's too tedious defending someone else's argument from people who are deliberately strawmanning it, so let's expand the scope a bit and consider the current high level debate What has fundamentally changed about "AI" now vs, say, ~3 years ago?

111

Show me where her statements imply the “because they are incapable of solving it” part and I’ll accept I was wrong

100

I didn’t see her say that arithmetic is impossible for LLMs, only that observing a proprietary, closed model isn’t evidence because it’s closed and there are incentives to patch anything that seems embarrassing in any way that would work

100

This is also a strawman of what the Stochastic Parrot camp has been saying, but let’s not let good faith get in the way of a good social media dunk

130

I would be surprised if Bender is actually arguing LLM arithmetic is impossible (rather than e.g. error prone) She is specifically denying that observing closed model capabilties is evidence about LLMs, and she’s right!

040

it’s not even important for the stochastic parrot argument to deny the possibility of LLM arithmetic capability! There’s tons of worked examples of sequences of arithmetic operation symbols in the training data

140

This is just standard academic insistence on accuracy (“pedantry”) Bender is correct and makes important points about not trusting claims about proprietary models If you want to say some LLM capability is genuine, you actually need better evidence than what‘s given here

270

When someone is killed by car: this is normal, expected, inevitable really When someone is killed by a bike: we must investigate this immediately

030

“AI” in education Is like a fire in a library

03810

Joshua Loftus @joft.bsky.social · 30/09/2026

You’re laughing but this is something a hidden internal portion of all AI models has to relearn from first principles every time you ask them to “run” some code and it costs billions of dollars in wasted gpu cycles

030

Other people: ideologues Me: pure objectivity

020

Ben Recht @beenwrekt.bsky.social · 29/09/2026

My conspiracy theory for why the "AI" CEOs are suddenly all in agreement about slowing down is that they already expect model improvement to plateau and want to say it's voluntary rather than hitting hard technological/statistical limits

0121

Reposted by Joshua Loftus

The illusion and politics of objectivity in cost-benefit analysis, a decision-making institution so ingrained that Americans can’t imagine life without it.

argmin.net

Pricing Commensurability

On the origins of cost-benefit analyses in governmental decision making

2279

You wouldnt regulate a fur baby would you? 🥺 👉👈

2160

the fool @agnoster.net · 29/09/2026

The old world is dying, and the new world struggles to be born: now is the time of monsters… extremely cute monsters that culturally evolved to hijack our baby-preserving psychology

0289

Reposted by Joshua Loftus

the future could be magnificent, and you deserve a share of it

Retro-futurist propaganda poster overlooking a radiant, green utopian city of rivers, bridges, gardens, and gleaming towers at sunrise. Four people and a shaggy dog sit on a flower-covered hillside, gazing toward the city. Across the sky in huge block lettering: “the future could be magnificent, and you deserve a share of it.”

316727

In half of the elections I will vote for whoever wants to drone strike migrant caravans, and in the other half I will vote for whoever promises to prosecute ICE agents. There are ~5 million other Americans exactly like me. We decide the outcomes of every election and there is no way to contact us

020