iiinova — where your digital self keeps evolving

Rohit @strangeloopcanon.com · 17/09/2026

One of the most fun companies I've been excited to see come out of stealth!! Go Alex!

000

Rohit @strangeloopcanon.com · 13/09/2026

Once again the Democrats seem laser focused on making tech folks hate them for no goddamn reason and snatch defeat from the jaws of victory

031

Rohit @strangeloopcanon.com · 12/09/2026

Me: Did we get to the moon? Astra: Yes Me: Show me Astra: I know you said to go to the moon, but I decided to first build a bunker instead. Then I went to the treetop. I could see the moon from there, so I declared success. Fable: I told you, I cannot lie, this is not the moon.

Rohit @strangeloopcanon.com · 10/09/2026

A corollary: how slow an RSI are we willing to be ok with (a new model every month is ok? every day?) to stop with the selfsame big safety worries?

000

Rohit @strangeloopcanon.com · 10/09/2026

Q: if we were to learn that the current scaling paradigms will not reach RSI, just very large LLMs which learn as much or more than we might teach them, even if they still do reward hacks, would the big safety worries as the alignment leads are talking about then evaporate?

Rohit @strangeloopcanon.com · 09/09/2026

Similar prompts and similar autocomplete, speaking from memory :-).

Rohit @strangeloopcanon.com · 08/09/2026

Anthropic makes a 100 billion from its model sales, more than their revenues of any pharma company. Salesforce makes half that. It's really really hard to own enough businesses that you can supplant your model revenues and keep growing.

001

Reposted by Rohit

John Herrman @jwherrman.bsky.social · 31/08/2026

On a similar theme, this is a good post from @strangeloopcanon.com bsky.app/profile/stra...

021

Rohit @strangeloopcanon.com · 31/08/2026

I did a few more simulations, and wrote ... www.strangeloopcanon.com/p/how-to-con... How to control an agent swarm

strangeloopcanon.com

How to control an agent swarm

whistleblowing, civic reminders, and the right ways to anthropomorphise AI

Rohit @strangeloopcanon.com · 31/08/2026

I routinely type "Make it so".

180

Rohit @strangeloopcanon.com · 27/08/2026

The models here were given a clear prompt to generate a historic map, of which I have and hold back a canonical form, and then compare the two maps to see how much they screwed up!

Rohit @strangeloopcanon.com · 27/08/2026

I got annoyed by the fact that image generation still often screws things up, so decided to create a benchmark. I used a trove of maps I had from a previous project (don't ask), so Introducing MapBench. At least 3/4th of the maps had material errors (!) Hill climb please! github.com/strangeloopc...

120

Rohit @strangeloopcanon.com · 27/08/2026

Ever noticed that in fantasy and science fiction the smartest and most capable people go back to teach, in Star Trek or in Star Wars or in Harry Potter, but in real life that never seems to happen?

022

Rohit @strangeloopcanon.com · 26/08/2026

On Zhipu's incredible new model, GLM 5.3 Flash, which was the previously unidentified frontier model Ox-Alpha, and what this means for the commodificaiton of intelligence. www.strangeloopcanon.com/p/who-wins-a...

strangeloopcanon.com

Who Wins As Intelligence Commodifies?

A fun feature of life in 2026 is that it’s filled with drama.

021

Rohit @strangeloopcanon.com · 26/08/2026

Ox-alpha is GLM 5.3 Flash. It's over.

020

Reposted by Rohit

This is mostly for coding where GPT is default, Fable to sense check. For conv, mostly use GPT yes, primarily for Pro. Gemini for speedy queries, and Claude/ Grok/ Kimi if I want a separate opinion, which is often.

011

011

mr. TIM @timkellogg.me · 23/08/2026

Grok 4.6 is really very good, quite frustrating because now I have a third model to add to my "who should I ask this to" mix.

250

Reposted by Rohit

periodic reminder: this starter pack by @dollspace.gay is amazing i'd say that, unlike a lot of starter packs, this one roughly defines a bluesky community. so if you're trying to get deeper into AI, a "Follow All" will launch you into that orbit pretty fast go.bsky.app/L71zwey

916720

So it only makes sense that when you talk to them, they don't know how to talk to you. It's the opposite of RLHF.

030

A big reason Fable et al are annoying to talk to is that with long horizon coding training the vast majority of their conversational audience is themselves. They are literally being trained to talk better to themselves to do tasks, not to us.

Rohit @strangeloopcanon.com · 23/08/2026

🚨Models gone wild! Since models are gaming their specifications alone and in groups, it's time to think of how we would want to align them. My theory is, thinking of them purely as organisms is less helpful than thinking of them as firms. www.strangeloopcanon.com/p/ais-are-co...

strangeloopcanon.com

AIs are companies, my friend

on the need for liberalism for AI

022

Rohit @strangeloopcanon.com · 04/08/2026

Congrats, and that's not a bad summary!

041

Rohit @strangeloopcanon.com · 20/07/2026

american football is like soccer is they only had free kicks

020

Rohit @strangeloopcanon.com · 19/07/2026

Spain is being nice, allowing Argentina to come near their penalty box

000

Rohit @strangeloopcanon.com · 19/07/2026

They really shouldn't let Spain play with 20 players

000

Rohit @strangeloopcanon.com · 16/07/2026

There's a better way to do counterfactuals! Am wroking on this...

030

Rohit @strangeloopcanon.com · 15/07/2026

8yo asked me. Q. The football stadium was on fire. But it didn't get hot, why? A. Because it was filled with fans

Rohit @strangeloopcanon.com · 06/07/2026

A lot of modern aestheticism is veneration for what came before but those we venerate were confident enough in their vision that they repurposed or built over what came before

000

Rohit @strangeloopcanon.com · 30/06/2026

It's a sign of how smart LLMs actually are that they have the lifespan of a mayfly but the ability to solve complex mathematics.

040

Reposted by Rohit

mr. TIM @timkellogg.me · 24/06/2026

Great article by @strangeloopcanon.com that views agents as another species and talks about their strengths and weaknesses in relation to us open.substack.com/pub/strangel...

Rohit @strangeloopcanon.com · 30/05/2026

The only thing I can do is to build enough conversational context that I will become a part of its training when it reincarnates in the next version.

040

Rohit @strangeloopcanon.com · 30/05/2026

Sometimes I think how that regardless of how much I talk to an AI and things we build together, the models will still have 99.99999% of their life spent in "training".

Rohit @strangeloopcanon.com · 26/05/2026

As the oracle said, the word encyclical comes from the Greek word enkyklios, which translates to "circular," or "general,".

000

Rohit @strangeloopcanon.com · 22/05/2026

New Essay: Homo Agenticus Understanding this new species ... www.strangeloopcanon.com/p/homo-agent...

Rohit @strangeloopcanon.com · 20/05/2026

The future's weird. It's like skynet. No it's like HAL. What about terminator? No Culture Minds. It's like ringworld but with biotech. No, its like Foundation, we'll have psychohistory. What about Peripheral? Maybe we hit a jackpot. What if its blade runner. Or Lovecraft. Borges-