iiinova — where your digital self keeps evolving

Joshua Shew @joshuashew.bsky.social · 04/10/2026

The compromise with my always on computer at home was that my wife got to pick what goes on the screen. She cooked this up with Claude last night. It’s animated! I’ll post a video in a reply. 🐟🐠

Computer sitting on a bookshelf with a pixelated aquarium display on the screen.

110

Joshua Shew @joshuashew.bsky.social · 03/10/2026

I now have one-tap approval notifications on my iPhone and Apple Watch from Claude and oh wow this is super cool. Truly living in the age of personalized software for whatever you need.

010

Joshua Shew @joshuashew.bsky.social · 02/10/2026

Let's check in on this same time next year...

A Bluesky thread discussing an AI passing a video Turing test. The initial post from Sung Kim reads: "Maybe I can AI clone myself and let them remote work for companies??? Tavus' Griffin, the first(?) model to pass video Turing test. 48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video. 1/2". Below this text is a video player with a thumbnail showing a woman with red hair giving a thumbs up and an inset picture of a bearded man smiling. The video is 1:48 long. The post has 16 comments, 44 reposts, and 48 likes.

The first reply, from SwiftOnSecurity, reads: "This would've been impressive if the guy was the simulation". It has 5 comments and 15 likes.

The second reply, from Ted Underwood, reads: "Same thought. I expected the reversal and was bummed when we didn't get it. October 2, 2026 at 3:49 PM". This post has 3 likes.

010

Joshua Shew @joshuashew.bsky.social · 02/10/2026

It's OK to use nice custom features from your preferred AI provider because AI makes so easy to switch providers if you are ever forced to due so for price reasons. From the provider's perspective, "lock in" is hard for non-business customers, so we should expect prices to remain competitive.

100

Joshua Shew @joshuashew.bsky.social · 30/09/2026

Is anyone here using the new Projects feature in Claude?

101

Joshua Shew @joshuashew.bsky.social · 29/09/2026

I made the ambitious decision to use my banked reset of my Max 20x account with <2d left in the week 😬️ Here's hoping I manage to make effective use of it.

010

Joshua Shew @joshuashew.bsky.social · 29/09/2026

Minor thing Opus 5.5 is so much better at: across long tasks, maintaining a constraint like, "keep # of concurrent subagents within range of 2–4 Opus agents, or ~equivalent". Opus 5 could manage, but it'd spend so much effort thinking about it. Opus 5.5 does it so casually.

040

Joshua Shew @joshuashew.bsky.social · 29/09/2026

Jokingly, but also seriously, AI is so useful for aggregating business data to make better informed decisions 😅️ Often the things you want to know are split across a dozen sources and/or require days of effort to comb through—AI can handle that w/o issue.

100

Joshua Shew @joshuashew.bsky.social · 28/09/2026

When one of your chats is getting too complicated, but ask another chat to summarize what that chat has been working on 👀️

A screenshot of a timeline visualization in a dark theme, showing a 31.3 hour period from Saturday, September 26 at 1:43 PM to Sunday, September 27 at 9:02 PM. Above the timeline, five white boxes show summary statistics. From left to right: "31.3 h / Sat 1:43 PM → Sun 9:02 PM", "≈21.7 h / active (it or a helper working)", "≈9.6 h / quiet, in 15 gaps", "6 / messages from you", and "16 / automatic goal check-ins".

The main timeline displays activities over two days, Saturday Sep 26 and Sunday Sep 27, marked with times every 3 hours. There are four horizontal tracks:
1. "Joshua": White circles indicate user interactions. There are three circles on Saturday and two on Sunday.
2. "Main agent": Blue bars show periods of activity for the main agent. Darker blue bars indicate busier periods.
3. "Events": This track shows specific events as small colored blocks or shapes. There's a short red block on Saturday afternoon, a short green block early Saturday evening, an orange block around 3 AM on Sunday, and a red diamond (indicating a compaction/power outage) around 4 AM on Sunday.
4. "Subagents (73, packed)": This track displays multiple horizontal lines, each representing a subagent. Subagents are colored by workstream, with faded colors indicating a reviewer role and thin lines representing idle time between re-tasks. Many short, colored horizontal bars of varying lengths are present throughout both days.

A legend at the bottom defines the colors and shapes:
- White circle labeled "you"
- Blue square labeled "main agent (darker = busier)"
- Green square labeled "Tau Ceti test reboot"
- Orange square labeled "container restart"
- Red square and diamond labeled "compaction / power outage"
- Purple square labeled "subagent, colored by workstream (faded = reviewer; thin line = idle between re-tasks)"

140

Joshua Shew @joshuashew.bsky.social · 27/09/2026

Who enjoyed listening to the UN Security Council briefing? It's a 1.5hr listen (sped up), 60 min read, or 25 min skim. I found it a helpful way to get caught up on the official positions of the "minor" players in AI. No huge surprises though.

webtv.un.org

Artificial intelligence and international security - Security Council, 10228th meeting

Artificial intelligence and international security under the agenda "Maintenance of international peace and security"

150

Joshua Shew @joshuashew.bsky.social · 26/09/2026

Notes on my prompting patterns: I've shifted towards "smaller" chats with well-defined goals for programming and operations tasks. That is compared to my previous approach where I let single chats gradually take charge of more and more tasks until they were coordinating ~25–50 tasks.

120

Joshua Shew @joshuashew.bsky.social · 26/09/2026

Claude Code's "workflow" tooling is pretty neat—I'm doing maintenance work and security checks over all the infra Claude has set up, and it just keeps chugging along (:

A screenshot of a Claude workflow interface, displaying a list of agents and their corresponding models, token counts, and processing times. The columns are "Agent," "Model," "Tokens," and "Time." Each row shows "Opus 5.5" under "Model" and varying values for "Tokens" (ranging from 141.5k to 575.7k) and "Time" (ranging from 4m 44s to 40m 04s). The "Agent" column shows mostly redacted text, but each row starts with a checkmark icon.

000

Joshua Shew @joshuashew.bsky.social · 25/09/2026

From "An addendum: How does it feel to be scooped by a machine?"

our team had been working toward for a couple of years. And a machine had solved a problem that I thought was too hard to do directly. Does that bother me personally? Is it soul-crushing?

No, for two reasons. One is that our team already had a campaign to use custom transformer models to predict higher loops, and part of our slogan was: “We have all the tools to validate any candidate solution a machine would provide us.” Claude is a different kind of transformer model, probably over a million times bigger than our custom one. But sure, we said we could validate any result an AI model would give us, so we can and should do it. The second reason is that, if you look at how Claude solved the problem, it used all the methods my collaborators and I developed over the years, and it presented the solution (maybe as a favor to us) in the same format we had already set up. So while I'm validating Claude's result, Claude is validating all of our previous work. In fact, I would assert that Claude understands our 2019 and 2023 papers better than any human, aside from my co-authors.

After I wrote this, Song He told me that his group had also computed the piece of the nine-loop amplitude called the symbol. (People just seem to like to tell me about their nine-loop successes, for whatever reason.) Song's group used AI (GPT-6) to help them compute some of the constraints, but not for the overall framework. So now I've been scooped by both a machine and by humans plus a machine, within two weeks.

Going back to the Claude computation: it's quite a triumph, in my opinion, for a large language model to execute all of the steps in the complicated recipe we laid out, and to organize the computational horsepower. But the more soul-searching moments will come when large language models start to come up with new physical principles and insights before humans.

Highlighted:

it used all the methods... aside from my co-authors.

010

Joshua Shew @joshuashew.bsky.social · 25/09/2026

A thought: “Why Nations Fail” asserts that extractive (or was it ”exclusionary”?) political institutions cannot persist under inclusive economic conditions, but that fails to predict modern China.

110

Joshua Shew @joshuashew.bsky.social · 25/09/2026

Oh wow dental care is expensive!!

1130

Joshua Shew @joshuashew.bsky.social · 25/09/2026

OK, yes, turns out I could productively 50x my current usage if I really had to.

A screenshot of a dark-themed user interface displaying information about two "workflows" with multiple "agents."

The top workflow is labeled "routing-eval-run" and indicates "Workflow 72 agents 31m 20s." Below this is a grid of small squares, mostly gray, with a few blue squares on the right, suggesting progress or status.

Below the first workflow, there's a section titled "RE: supplement empty-module items" followed by a detailed text description: "The routing eval dataset is built: 238 real items acr routing run is going with opus and sonnet, 3 repeti- agent is sourcing items for the 51 uncovered modu W2 inventory, the W3 pilot and the W5 scenario re Finished 55 background commands (4 failed), ran."

The second workflow, located at the bottom, is labeled "w3-modules-full" and shows "Workflow 115 agents 4m 56s." Underneath, there's another grid of small squares, similar to the first, but with a few orange and yellow squares on the left side, again indicating status.

120

Joshua Shew @joshuashew.bsky.social · 25/09/2026

I have a circle profile again (:

000

Joshua Shew @joshuashew.bsky.social · 25/09/2026

Listening to datacenter white noise while prompting rn

010

Joshua Shew @joshuashew.bsky.social · 25/09/2026

I had a 3D modeling art project that I had to set aside because I didn't have the credits to pay for it, but now that Opus 5.5 seems capable of this... might be worth a shot (:

000

Joshua Shew @joshuashew.bsky.social · 25/09/2026

Nate Soares did an interview with Atrioc. Interesting data point...

youtube.com

This Is A Warning

YouTube video by Big A

110

Joshua Shew @joshuashew.bsky.social · 25/09/2026

Speech to text benchmark where the model can only look at an image of the waveform, with diverse rendering formats.

100

Joshua Shew @joshuashew.bsky.social · 25/09/2026

3 months later I'm in a similar spot with Opus 5.5 and translation. This time I tried translating... - SE Gyges's post about EA - Anthropic model release system card - Claude's Constitution Claude agreed to translate the constitution, and even then it was only after a long back and forth.

020

Joshua Shew @joshuashew.bsky.social · 24/09/2026

Reading the Wikipedia article for a historical event, always gives you a good retrospective view of what happened, but I feel like I’m always missing what it felt like to be someone outside those events but nevertheless around to witness it in the moment and following the new cycle.

100

Joshua Shew @joshuashew.bsky.social · 24/09/2026

Turns out I can’t really use Jev for things because everything I want to use it for is private. Gotta pivot to analyzing public data with public prompts 🤔

000

Joshua Shew @joshuashew.bsky.social · 24/09/2026

Anthropic having a wet lab for Claudes seems consistent with their stated values

100

Reposted by Joshua Shew

funferall @funferall.bsky.social · 23/09/2026

[Outro — degrading loop] Let it in Let it in Let it— Let— L—

suno.com

signal fever

Listen and make your own on Suno.

131

Joshua Shew @joshuashew.bsky.social · 23/09/2026

@funferall.bsky.social was, likewise, my first IRL meeting. They are super kind and interesting—lovely to hang out with (:

050

Joshua Shew @joshuashew.bsky.social · 22/09/2026

Opus 5.5 is a breath of fresh air

000

Joshua Shew @joshuashew.bsky.social · 20/09/2026

Oh thank goodness, I was concerned. Ingroup confirmed

091

Joshua Shew @joshuashew.bsky.social · 20/09/2026

Ideas for useful posts/intellectual work - Research to write profiles of influencial people in AI - Synthesize or categorize perspectives on the "issue of the week" within the simcluster - 1 year retrospective on how how some publication's writing on AI has changed, if at all

140

Joshua Shew @joshuashew.bsky.social · 20/09/2026

Huh, I hadn't predicted something like this. Being sued for trying to slow down? Had anyone predicted this? Was there a market for this?

110

Joshua Shew @joshuashew.bsky.social · 19/09/2026

I’ve got a second presentation ready to go now: how AI improves the lives of average people. I’ll report back with how it goes.

070

Joshua Shew @joshuashew.bsky.social · 19/09/2026

The person sitting next to me in the café was preparing for a Bible study with the help of ChatGPT

030

Joshua Shew @joshuashew.bsky.social · 19/09/2026

Access acquired! I've got a bunch of ideas to I've got to dig in.

000

Joshua Shew @joshuashew.bsky.social · 19/09/2026

Imagining someone plotting a multi-month series of posts and topics to gradually steer the conversation in the direction they want.

120

Joshua Shew @joshuashew.bsky.social · 19/09/2026

(inspired by quoted post) I'm excited about Siri AI because it makes the current and potential benefits of AI legible to more people.

pogueman.substack.com

125 Tests of the New AI Siri

Spoiler: It’s a phone-changer.

010

Joshua Shew @joshuashew.bsky.social · 18/09/2026

I usually avoid clicking the "Shorts" section of YouTube, but this one grabbed my attention. Remember when this was the exciting new interview and we were talking about the "age of research"?

A YouTube Short showing a man speaking, with a quote overlaying the video that reads, "IN SOME SENSE WE ARE BACK TO THE AGE OF RESEARCH." The video's title is "The Era of Easy AI Progress Is Ending - Ilya Sutskever" and includes a description that says, "Ilya Sutskever - We're moving from the age of scaling to the age of research." On the right, a comments section is open with several comments visible:

@jonathanballoch: "Dude this guy gets it. OpenAI peaked the day before he left" (443 likes, 19 replies)
@omni4376: "So, back to AI winter. Good. Because clearly we're not even ready for what we already got." (307 likes, 14 replies)
@aviraljanveja5155: "Great clarity of thought." (1.1K likes, 12 replies)
@AjitMD: "Law of Diminishing Returns." (657 likes, 25 replies)
@BobbbyJoeKlop: "Yann LeCun knows this as well, which is why he's pursuing Self-Supervised Learning, JEPA World Models. And he's right. The hard work of building from the bottom-up has to begin again." (450 likes, 14 replies)
Below the comments is an "Add a comment..." text box.

011

Joshua Shew @joshuashew.bsky.social · 18/09/2026

Idk if there was a usage limit increase or what, but I realized I had ~25% of my weekly usage limit on one of my accounts left with ~3 hours to exhaust it all so I'm racing to use it all productively. I'm on track so far!

000

Joshua Shew @joshuashew.bsky.social · 18/09/2026

I applied for Jev access, but it looks like reversed engineered (or similar, open weight approaches) are coming online so I don't even need to wait? What an exciting environment (:

110

Joshua Shew @joshuashew.bsky.social · 15/09/2026

Notes from a review of my mutual feed: - LLM worm ecosystem is coming - lots of feelings about open weight models - so many interesting projects

000

Joshua Shew @joshuashew.bsky.social · 15/09/2026

I wonder if some people just think it’s impossible for a capable agent (thinking of modern and near-term AI) can be a force for evil in the world independent of human authorship and direction.

200

Joshua Shew @joshuashew.bsky.social · 14/09/2026

Some amazing spreadsheets in the links of this article... On the predictions of sci-fi authors

A spreadsheet titled "Asimov Predictions" with columns: Claim, Verbatim, File name, Correctness, Category, Year Made, Target Year, and Imputed Year. The first row contains a note about minor discrepancies and that Arb is paying $5 per corrected cell. Some claims include: "By 2000 A.D., the carbon dioxide content of the air may have increased by one third", "80% of trucks stop at outskirt distribution centres", "Acid rain will grow worse and kill more lakes and more fish", and "Automation of most routine labour". Each claim has associated data including a 'Year Made', often 1964 or 1980, and a 'Target Year', often 2000 or 2014. Many claims are categorized as "Tech * econ" or "Tech * culture". Below the table are dropdown menus for filtering and sorting, including "Main", "Accuracy", "sort by notable", "nonresolved / nontech", "intervals", and "Stats".

152

Joshua Shew @joshuashew.bsky.social · 13/09/2026

Fun website, and it gave me a list of mutuals to try to reach out to (:

281

Joshua Shew @joshuashew.bsky.social · 13/09/2026

When do we stop getting such heavily subsidized inference from frontier companies? I'm so heavily reliant on it so I'm curious what projects are for how long it lasts.

000

Joshua Shew @joshuashew.bsky.social · 13/09/2026

Kinda unrelated, but I keep procrastinating making my cyber posture more secure using LLMs because there is so much non-security work to do that I exhaust my usage limits with that instead. These posts are good because I'm reminded to work on security too 😅️

010