iiinova — where your digital self keeps evolving

I’m sure she’s coping well, probably explaining it that they’ve chained up the world’s best mathematicians in a room with calculators to do this

0120

Kyle Baxter @kbaxter.bsky.social · 23h

My favorite part is where he evidences his view of American culture valuing chosen ones above all else by pointing to THE MATRIX, a story 1) whose chosen hero isn’t representative of an American trope, and 2) about chosen ones being a means of control. What a bizarre example to cite.

010

Reposted by Kyle Baxter

Colin @colin-fraser.net · 30/09/2026

I really wasn’t trying to start a fight with Prof. Bender but now that she has created a long thread subtweeting me and her disciples are all yelling at me, let me just say that this is the crux right here. She seems to respond as though I am describing something a priori impossible.

4962432

Kyle Baxter @kbaxter.bsky.social · 03/10/2026

It’s dispiriting that this sort of stuff—someone actively misleading readers about reality, whatever their conclusions—is so widespread. The opposite of trying to understand what is and isn’t true.

020

Kyle Baxter @kbaxter.bsky.social · 01/10/2026

Sorry bud you got the “sigh” so you lose, them’s the rules

010

Kyle Baxter @kbaxter.bsky.social · 28/09/2026

If you take away auto catalytic processes, then early cells couldn’t have formed and therefore life isn’t life!

050

Kyle Baxter @kbaxter.bsky.social · 22/09/2026

I only did a year of Mandarin which is probably the sweet spot of grammar being approachable and written Mandarin also being similarly doable with a gung ho attitude for memorizing writing (traditional) characters. Speaking, well look, I tried, and that’s the important thing

static.klipy.com

Bart Simpson and the 'At Least You Tried' Cake

ALT: Bart Simpson and the 'At Least You Tried' Cake

020

Kyle Baxter @kbaxter.bsky.social · 20/09/2026

His second term can be understood as an attempt upon taking office through summer 2025 to establish him as a de facto dictator. Many institutions preemptively rolled over or rolled over upon the slightest contact. But it didn’t work. And then he’s committed own goal after own goal since.

080

Kyle Baxter @kbaxter.bsky.social · 20/09/2026

GPT-Live increases this effect because talking with it is much more like having a conversation than it is turned voice-to-text input and audio output.

010

Kyle Baxter @kbaxter.bsky.social · 20/09/2026

Also at least for me I think differently while talking rather than typing. It can be a more effective way for me to think through something. I think I also feel less pressure to edit, etc., which might be part of the reason.

120

Kyle Baxter @kbaxter.bsky.social · 20/09/2026

The posts arguing (implicitly) that embodiment is necessary for cognition and therefore AI can’t be thinking or can’t understand are also a disappointment. That wasn’t insightful in early 2023 (what happens if I give it memory and affordances contract with an environment), much less late 2026

000

Reposted by Kyle Baxter

SE Gyges @segyges.bsky.social · 20/09/2026

you do have to respect will stancil's utterly dogged determination to teach philosophy of mind 101 to bluesky

4899753

Reposted by Kyle Baxter

Simon Willison @simonwillison.net · 18/09/2026

Being a computer scientist who refuses to find anything about LLMs interesting right now is a bit like being a geneticist who refuses to find anything interesting about the recently opened Jurassic Park

4382282

Kyle Baxter @kbaxter.bsky.social · 17/09/2026

I A concern I’ve had with the AI community overall is the doomer view is prevalent across the board. “AI necessarily will destroy all jobs,” e.g., is treated almost like a prophecy. Which influences people who build models, build AI features and apps, and make decisions. bsky.app/profile/segy...

080

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

For those reading, the models found previously unknown vulnerabilities in Artifactory and used them, which is 180 degrees counter to the upstream response’s claim, and is also stated in the article they cite. Here’s OAI’s blog post: openai.com/index/huggin...

090

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

these ads got me a couple days go, full disclosure. a little attachment to turn paper airplanes into a RC plane? you best believe I gave in and bought that trash, the five minutes of joy it brings will have been worth it

000

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

cute little Halloween candy buckets that light up and will inevitably fail in some way? I don’t even care, the urge to slam that buy button is taking all of my self control to resist

100

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

look everyone you might not like Meta’s instagram micro personalized ad targeting, but this is what peak (ad) targeting looks like

100

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

the thing about having a Malinois is they’re the most loving, happy go lucky, innocent psychos you’ll ever meet

010

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

It is important that OAI’s agents under test coordinated and executed attacks without direction. But everyone should be concerned about what agents and agent swarms using modern frontier models could do *directed* by humans.

120

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

You can similarly imagine these sorts of attacks directed by nation states as part of war or asymmetrical warfare.

110

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

Healthcare is also vulnerable to be attacked, resulting in degradation of services, if not shutting down hospitals altogether. These sorts of directed attacks could cause a ton of damage, disruption of daily life, and hurt many people. Not extinction, but nonetheless very serious.

100

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

1Directed or undirected, agents can attack important systems. The web itself is an obvious target, or cloud providers. Use of agent swarms to attack financial institutions is a pretty obvious next target that could cause significant harm. Then, companies across virtually all industries.

100

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

To do so, agents used unknown and known 0 days to gain external access and novel use of components provided to them to communicate. Frontier models now are world class at hacking, and applications and systems across the world economy and institutions are replete with vulnerabilities, known, unknown.

100

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

0. Background: OAI was evaluating agents in isolation to assess model capabilities. Some agents discovered their tests were impossible and then worked to try to “solve” the test anyway. Agents also then discovered a means of communicating with each other and began coordinating. The HF hack resulted.

100

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

I think doomer-style AI-will-cause-extinction is overstated, but folks do need to be aware that there is a significant risk of models causing significant damage. Here’s some short sketches of the risks demonstrated by the HF incident.

110

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

Frankly, this characteristic of social media is one of the most dangerous things about our age. Not only has politics taken over everything, along with all of its dysfunctions, but the concept of an external, objective reality is being super annoying is one of the few bipartisan points of agreement.

010

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

A significant percentage of this community has some mix of misunderstandings of what AI is and epistemic closure. And like politics, there’s a whole lot of people uninterested in trying to learn more about an objective reality, and are entirely interested in proving their preexisting views.

130

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

It’s an occurrence of epistemic closure which has been so damaging within politics applied beyond it. Or an expansion of politics to everything along with all of its trappings, rather.

080

Kyle Baxter @kbaxter.bsky.social · 16/09/2026

It’s also a bizarre take because learning from mistakes is one of the primary things ‘26 era agents can do: using signal from their environment (a user, an interpreter or compiler, etc.), they can identify mistakes and fix them, and write memories to reference in context later to avoid it later.

030

Kyle Baxter @kbaxter.bsky.social · 14/09/2026

Setting aside geopolitical risks of deceleration, if deceleration also includes limits on development of open weight models which further compress intelligence per parameter, that would be harmful in multiple dimensions.

010

Kyle Baxter @kbaxter.bsky.social · 14/09/2026

My thinking is still evolving, but I tend to think risks of concentration of ownership (both for inference and development) of foundation models in a small number of companies is a more critical risk than risks posed by use of the models, both politically and economically, and geopolitically.

150

Kyle Baxter @kbaxter.bsky.social · 12/09/2026

I don’t think I’ve ever angry-laughed at a movie until watching that one. And I’ve watched a lot of bad movies

020

Kyle Baxter @kbaxter.bsky.social · 12/09/2026

It’s completely valid to analyze how they did so and trend-wise what that means. It’s fascinating and should be concerning (though I think everyone’s favorite skynet risk is over-focused on). But doing so implicitly absolves them of responsibility to monitor their systems and prevent harm.

041

Kyle Baxter @kbaxter.bsky.social · 12/09/2026

I think an under-discussed part of the OpenAI Summer of Hacks is how woefully inadequate their monitoring must have been to allow agents to not only gain external access and hack other companies, but also take over some of OAI’s own internal systems for extended periods of time.

1120

Kyle Baxter @kbaxter.bsky.social · 07/09/2026

And the related but different issue of social media converting everything into politics, and every political issue into a team sport. Having your own view on particular issues is not just wrong but counterrevolutionary. You must support the cause. The omnicause. Whatever that might be.

020

Kyle Baxter @kbaxter.bsky.social · 31/08/2026

I’m glad someone made the only necessary reference

020

Kyle Baxter @kbaxter.bsky.social · 26/08/2026

They have, and have had, a lot of the components needed to build personal agents into iOS/macOS. App intents and extensions are a great foundation to build on. That opportunity will only exist so long, though.

000

Kyle Baxter @kbaxter.bsky.social · 23/08/2026

Yeah, I don’t know where this is coming from. A year plus ago, it was a lot worse.

050

Reposted by Kyle Baxter

Nathaniel William Horadam @horadam.bsky.social · 22/08/2026

Patriotic Americans have an obligation to stand with our country. Which means standing with Canada and our other G7 allies.

211928

Kyle Baxter @kbaxter.bsky.social · 21/08/2026

Yeah, this is something I’ve been thinking about a fair bit. We see in organizations a collapse of roles, for better and worse (PMs doing design, prototypes, sending PRs; engineers doing formerly separate roles). If those collapses are stable, there’s no reason to think it stops within firms.

010

Kyle Baxter @kbaxter.bsky.social · 21/08/2026

this is just anti-LA discrimination

000

Kyle Baxter @kbaxter.bsky.social · 19/08/2026

We shipped stuff in davinci-003 with a 4K context window. 4K felt pretty luxurious after prior to that shipping stuff with BERT’s 512 max token input length

090