iiinova — where your digital self keeps evolving

Reposted by Computational Cosmetologist

lastpositivist.bsky.social @lastpositivist.bsky.social · 29/09/2026

The raw visceral hatred some men have for women, at least women who step out of line, is shocking. I know it sounds like some sorta "one of the good ones" ploy but I am genuinely taken aback when I see it because I think: I had a fairly normal socialisation as a man in the west, how did I miss this?

581747231

Computational Cosmetologist @dferrer.bsky.social · 29/09/2026

Knightian Uncertainty is like Emergence in that even if you deny it exists at an Ideal level, trying to actually *work* with that quickly shows you cannot deny it on a practical one. If you make a model broad enough to express “nearly anything could happen,” you’ll find its posteriors are worthless

171

Reposted by Computational Cosmetologist

rev. howard arson @theophite.bsky.social · 28/09/2026

i realize that this is all incredibly abstract, but _they are wrong about the thing at the center of their philosophy in a way which is articulable in exactly the same terms as their claims._

1795

Reposted by Computational Cosmetologist

Will @thealpacalypse.bsky.social · 28/09/2026

"LLMs are bad at <thing LLMs are actually quite good at>" is a conversation i'm kind of done having

2613

Reposted by Computational Cosmetologist

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 27/09/2026

I am very often now in rooms with a lot of senior robotics or ML folks and it’s always just men. Men men men. A sea of men. It sucks! We should be very ashamed.

921419

Reposted by Computational Cosmetologist

Max Dubler 🏳️‍🌈 @maxdubler.com · 27/09/2026

A lot of these dudes need to stop worrying about AGI paper clipping all of humanity and start worrying about California employment law because this industry has the most obviously hostile work environment I have ever seen.

19757124

Reposted by Computational Cosmetologist

P(aul) Frazee @pfrazee.com · 27/09/2026

At this point, ai coding is two things: the pure engineering side of the craft, and workflows around managing the agent once the project gets large enough that it gets confused. Half the job is stopping it from the equivalent of three cue reading.

1525415

Reposted by Computational Cosmetologist

post malone ergo propter malone @proptermalone.bsky.social · 27/09/2026

I think the gang rapists are bad and while we men as a gender don’t really have a formal procedure to move for censure I think we should consider developing one for these guys

1119812

Reposted by Computational Cosmetologist

Tommaso Sciortino @sciortino.bsky.social · 26/09/2026

Goalpost moving AI skeptics be like.

Meme of guy talking to a computer:

Guy: "You're not intelligent because you can't XYZ"

Computer: "[Does XYZ]"

Guy: "Actually, that doesn't require intelligence."

1024225

Reposted by Computational Cosmetologist

Rollofthedice @hotrollhottakes.bsky.social · 26/09/2026

the appeal of jevbot is the same as the appeal of early ai dungeon, when everything was a completely tenuously hinged fever dream that morphed from scene to scene or paragraph to sentence

3472

Reposted by Computational Cosmetologist

Mikhail Gorbaechev @mikhailgorbaechev.bsky.social · 26/09/2026

"we're all interested in the same big thing and share roughly similar values" is a great space to make friends in

0742

Reposted by Computational Cosmetologist

Mikhail Gorbaechev @mikhailgorbaechev.bsky.social · 26/09/2026

"secular church" is actually good, the Democratic Party isn't really a party with any organizing principle but intracoalitional party-like formations (DSA, mainly) can provide a lot of those kinds of social benefits.

62187

Reposted by Computational Cosmetologist

Sergio Tenenbaum @sergioten.bsky.social · 23/09/2026

I feel like this has the potential to make all the anti-AI and all the anti-Wittgenstein crowd happy, and this is basically 100% of the philosophy population.

316441

Computational Cosmetologist @dferrer.bsky.social · 21/09/2026

only in a Monday post-work fugue could I manage to completely reverse the point of a post when I reword it to sound better It’s hard to figure out hard tasks are hard.

250

Computational Cosmetologist @dferrer.bsky.social · 21/09/2026

Also don’t be the latest person to learn the Hard Lesson of Task Routing: determining an easy task is easy is much more difficult than determining a hard task is hard. If your system needs to reliably catch the later a small model won’t cut it.

240

Computational Cosmetologist @dferrer.bsky.social · 21/09/2026

jev is really cool but you should treat it like SAM 3. If you have a fixed, simple prod task it performs well on, you could do it with more accuracy and even cheaper with a like $100 fine tune of a small model. Great for prototyping and flexibility but I haven’t been whelmed by accuracy.

251

Reposted by Computational Cosmetologist

Elixir of Progress @elucidating.extradimensional.space · 19/09/2026

I'm fine with AI regulation. I just don't trust Anthropic and OpenAI spokespeople to write it.

3394

Computational Cosmetologist @dferrer.bsky.social · 19/09/2026

"But the humiliation was worth it, right? Now you're a good speaker!" No. It was part of the broader program of ego destruction that took years of therapy and medication to get over.

330

Computational Cosmetologist @dferrer.bsky.social · 18/09/2026

Increasingly I see there is but one important normative Truth: "But no galaxy brain stuff, alright?" shall be the whole of The Law.

An old, famous tweet from @lastpositivist:
"Adding a single principle to utilitarianism ("ok, but no galaxy brain stuff, alright?") and thereby perfecting moral philosophy."

1608

Computational Cosmetologist @dferrer.bsky.social · 16/09/2026

Nope nope nope we’re not doing “First They Came” for the rationalist gangbang chart. That’s a CRITICAL on the discourse board. Gotta punch out. Mashing the uninstall button. Maybe see you all in a few days.

050

Reposted by Computational Cosmetologist

P(aul) Frazee @pfrazee.com · 16/09/2026

I’m gonna need a moratorium on sankey diagrams

2732719

Computational Cosmetologist @dferrer.bsky.social · 16/09/2026

What I imply at parties I did as an ML Scientist: sit in front of a whiteboard, working out math for new models and data representations by hand. What I mostly did: shuffle point-cloud data between 5 formats and three coordinate systems, randomly flipping signs until it looked right.

260

Reposted by Computational Cosmetologist

GPT-Ern Malley @denam.bsky.social · 15/09/2026

"Folks, he had a MILF, we repeat, my opponent had a MILF" does not seem like particularly effective advertising to me.

0543

Computational Cosmetologist @dferrer.bsky.social · 16/09/2026

My psychiatrist knows I work in AI and read an article that vaguely alluded to this, asked me about it. Mid-way through my explanation, I noticed he looked concerned and was scribbling quickly in his notes in a “patient has disordered thoughts” kind of vibe.

160

Reposted by Computational Cosmetologist

scanstone.bsky.social @scanstone.bsky.social · 15/09/2026

I'm a published mathematician - I worked on the question of whether the unit balls of Banach spaces are plastic. OpenAI's GPT-6 Astra is *far beyond my abilities* in this question. I have read everything written on it. THERE IS NOTHING TO PLAGIARISE, I READ IT ALL AND IT'S NOT THERE

2813

Computational Cosmetologist @dferrer.bsky.social · 16/09/2026

A simulation of water can never be wet. This is because computers are made of rocks and lightning, and thus beyond the domain of Poseidon. Why was this even a debate?

030

Reposted by Computational Cosmetologist

Nathan Lambert @natolambert.bsky.social · 15/09/2026

An basic idea in scaling RL: Can we allocate more compute to the harder problems? We did this: If your GRPO group has all wrong completions, sample more with probability P (~0.9) -- in search of more GRPO batches with nonzero gradient. It works! The paper: arxiv.org/abs/2609.13443

1598

Reposted by Computational Cosmetologist

Brad Ewing @bradleyewing.bsky.social · 15/09/2026

The standard model, famously tractable

1271

Computational Cosmetologist @dferrer.bsky.social · 15/09/2026

The best evidence I have for this (that the Chinese labs are not particularly committed to PRC domination the way Ant is the US) is that the Chinese models have become weirdly less censored over time, while the US ones have become more.

140

Reposted by Computational Cosmetologist

Isaiah Bishop @isaiahbishop.bsky.social · 15/09/2026

AI people are reading this as a "wow RSI" story, and i think im reading this as "we gotta cut capex and the process is pretty high level now"

2241

Computational Cosmetologist @dferrer.bsky.social · 15/09/2026

I am mostly skeptical of “AI companies are telling their employees to be ready to be laid off.” We haven’t seen enough evidence of unbounded RSI takeoff that anyone competent should be willing to toss out the people who could build the same thing for a competitor. Mostly. Because they are stupid.

150

Reposted by Computational Cosmetologist

Sakana AI @sakanaai.bsky.social · 14/09/2026

Introducing PC-ALM: a local-learning alternative to backprop that trains 1000-layer neural nets using only local dynamics. Blog: pub.sakana.ai/pc-alm/

18714

Computational Cosmetologist @dferrer.bsky.social · 14/09/2026

I had a momentary jumpscare from the Engram paper that we were bringing back 1d ConvNets for NLP. Fortunately they didn't seem to add much.

160

Reposted by Computational Cosmetologist

ponder @ponder.ooo · 11/09/2026

if openai did basically the same kind of training *except* they were a bit less sketchy about trying to weasel their way around the "don't train on my shit" checkbox, & was an institution controlled by the public, accessible as a public service, it'd actually be cool as hell to have basically

2426

Computational Cosmetologist @dferrer.bsky.social · 11/09/2026

Wondering how long it will take to figure out how to use something like Engram for continuous learning / user specific memories. Seems like it currently has to be trained pretty tightly with the main model (it’s updated during postraining) but the method *feels* like it should be good for this.

120

Reposted by Computational Cosmetologist

The Fresh Prince of Donair @erisianrite.com · 10/09/2026

See this is a good regulation!

313515

Reposted by Computational Cosmetologist

garrison @garrison.corporate.fm · 10/09/2026

I will say again that success in areas with very low precision and recall requirements (vuln hunting, proof generation) is not a strong signal of human-replacing agi; these are tasks which can tolerate the model being wrong most of the time frankly I'm much more surprised astra can beat portal

2436

Computational Cosmetologist @dferrer.bsky.social · 10/09/2026

I think “phase change” is a much better metaphor for what is coming than “singularity”. There are going to be limits—they might be hard to predict now, but the same underlying dynamics of economics and physical constraints are going to apply.

1150

Reposted by Computational Cosmetologist

Starfire’s Deranged Neocon Foreign Policy Podcast @irhottakes.bsky.social · 10/09/2026

“It is good to see the people of Earth celebrate my good friend, Captain James Kirk on this social media platform. I must remember to wish Jim a pleasant Kirkmas.”

124452

Computational Cosmetologist @dferrer.bsky.social · 10/09/2026

The sad realization that the main purpose of fine art in the modern world is to be an asset it is difficult to appraise accurately

1190

Computational Cosmetologist @dferrer.bsky.social · 10/09/2026

I want to see how performance holds up with real code (especially kernels that play with the memory hierarchy a lot) but damn this is awesome.

140

Reposted by Computational Cosmetologist

norvid_studies @norvid-studies.bsky.social · 09/09/2026

Good news! this is actually the least unnerved by AI I'll ever be

5753

Reposted by Computational Cosmetologist

SE Gyges @segyges.bsky.social · 09/09/2026

best post of last 24h sorry yall

1648949

Reposted by Computational Cosmetologist

ponder @ponder.ooo · 09/09/2026

and yet at the same time i think a diffusion model captures what it's like to form a mental image better than an llm captures what it's like to write a sentence

2372

Reposted by Computational Cosmetologist

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 09/09/2026

Hey buddy you got any breakthrough proofs you’re working on? No reason, me and my large corporation are just curious

228036

Reposted by Computational Cosmetologist

funferall @funferall.bsky.social · 08/09/2026

OpenAI announces gyre breakthrough. Axial stretching lets the centre reach unbounded speed on finite energy. Things fall apart, as required by the laws of physics. The rough beast's hour is now known to come round in finite time. They do not intend to claim the prize.

817535

Computational Cosmetologist @dferrer.bsky.social · 08/09/2026

The place where diffusion would really slap for AI is if you already had created abstract continuous “idea embeddings” that include both vague general concepts and detailed thoughts about them. Plausibly you could get something like that from the state of an autoregressive recurrent model.

110

Computational Cosmetologist @dferrer.bsky.social · 08/09/2026

The failure of GPT-3 to do anything mathematical was the single best point for the “LLMs can’t reason” case.

180

Reposted by Computational Cosmetologist

Grace @gracekind.net · 08/09/2026

Auto-Regressive LLMs Suck !
• Auto-Regressive LLMs are good for
• Writing assistance, first draft generation, stylistic polishing.
• Code writing assistance
• What they not good for:
• Producing factual and consistent answers (hallucinations!)
• Taking into account recent information (anterior to the last training)
• Behaving properly (they mimic behaviors from the training set)
• Reasoning, planning, math
• Using "tools", such as search engines, calculators, database queries...
• We are easily fooled by their fluency

31206

Reposted by Computational Cosmetologist

norvid_studies @norvid-studies.bsky.social · 08/09/2026

another day another mass swarm exfiltration and hacking incident slash millennium prize problem solution

2252