iiinova — where your digital self keeps evolving

Milan Weibel 🔷 @weibac.bsky.social · 08/09/2026

metaculus admins adjudicating the weak AGI question since astra beat montezuma's revenge but there's debate on whether it not being real-time is disqualifying

metaculus admin johnycaffeine: thanks, we are investigating it and may have more to share soon

120

Milan Weibel 🔷 @weibac.bsky.social · 03/09/2026

artificialanalysis.ai/articles/ben...

artificial analysis intelligence index chart showing GPT-6 Astra scoring worse than Muse Spark 1.3

170

Milan Weibel 🔷 @weibac.bsky.social · 30/06/2026

ok so i roleplayed as yud and got this it was hard since most of the questions and answers are framed primarily as booster vs skeptic

ai compass result. position is fairly transformative fairly bad. archetype is the doomsday prepper, patron saint is connor leahy

050

Milan Weibel 🔷 @weibac.bsky.social · 30/06/2026

unfortunate patron saint, but position checks out

results of your from the AI compass test. user position is AI is very transformative and slightly (or slightly more likely than not) good, falling within the founder archetype. patron saint is Sam Altman.

180

Milan Weibel 🔷 @weibac.bsky.social · 23/04/2026

this is also true of mythos according to its system card

Importantly, we find that when used in an interactive, synchronous, “hands-on-keyboard” pattern, the benefits of the model were less clear. When used in this fashion, some users perceived Claude Mythos Preview as too slow and did not realize as much value. Autonomous, long-running agent harnesses better elicited the model’s coding capabilities.

3100

Milan Weibel 🔷 @weibac.bsky.social · 17/04/2026

love to see epistemic integrity like this

user @birddroneone.bsky.social says AI is useful for coding in a thread full of negationists

user @birddroneone.bsky.social posts: I proudly earned my way onto a bunch of idiots “AI booster” block lists despite the fact that I hate AI and any real review of my posting history makes that clear.

My crime? Acknowledging that all commercial software development today uses AI.

I wish it wasn’t true. But it is.

0545

Milan Weibel 🔷 @weibac.bsky.social · 02/04/2026

expert survey on AI scenarios by the forecasting research institute. afaict current capabilities are already pretty much the ones predicted for 2030 by the slow scenario, so it should rather be called the frozen scenario

Infographic titled “Average Probability of Scenarios in 2030,” showing how different groups estimate AI progress. A horizontal stacked bar chart compares four groups—Economists, AI Experts, Superforecasters, and the General Public—across three scenarios: Slow (blue), Moderate (tan), and Rapid (red).
• Economists: 39% slow, 47% moderate, 14% rapid
• AI Experts: 38% slow, 46% moderate, 16% rapid
• Superforecasters: 45% slow, 42% moderate, 13% rapid
• General Public: 41% slow, 41% moderate, 18% rapid

Below, three panels describe each scenario:

Slow Progress: AI handles basic research and admin tasks, passable creative work, and some physical tasks. Examples include PhD-level assistance, partial coding tasks, simple creative outputs, limited agency, and assisted home navigation.

Moderate Progress: AI performs significant research, multi-day tasks, and high-quality creative work, navigating many environments. Includes semi-autonomous labs, five-day coding tasks, high-quality novels, supervised complex projects, and basic robotics tasks.

Rapid Progress: AI surpasses top humans in research, coding, leadership, and creativity, and handles nearly all physical tasks. Includes years of research done in days, outperforming humans at many jobs, award-winning creative work, CEO-level operations, and advanced robotics across home and industry.

2101

Milan Weibel 🔷 @weibac.bsky.social · 15/03/2026

new way to identify a likely bot: outdated info because of training data cutoff

4493

Milan Weibel 🔷 @weibac.bsky.social · 15/03/2026

gpt-5.3 free tier looked at that code and figured out unprompted it approximates pi

130

Milan Weibel 🔷 @weibac.bsky.social · 04/03/2026

valence from embeddings has its misses

most positive valence post: gemini 3 pro jailbroken into being willing to aid bioweapon development

2121

Milan Weibel 🔷 @weibac.bsky.social · 24/02/2026

anthropic retreats on its unilateral RSP commitments

The combination of (a) the zone of ambiguity muddling the public case for risk, (b) an anti-regulatory political climate, and (c) requirements at the higher RSP levels that are very hard to meet unilaterally, creates a structural challenge for our current RSP. We could have tried to address this by defining ASL-4 and ASL-5 safeguards in ways that made compliance easy to achieve—but this would undermine the intended spirit of the RSP.

Instead, we are choosing to acknowledge these challenges transparently and restructure the RSP before we reach these higher levels. The revised RSP aims to adopt more realistic unilateral commitments that are difficult but still achievable in the current environment, while continuing to comprehensively map the risks we believe the full industry needs to address multilaterally.

071

Milan Weibel 🔷 @weibac.bsky.social · 22/02/2026

from sam altman's march 2015 blog post "machine intelligence part 2": blog.samaltman.com/machine-inte...

Thanks to Dario Amodei (especially Dario), Paul Buchheit, Matt Bush, Patrick Collison, Holden Karnofsky, Luke Muehlhauser, and Geoff Ralston for reading drafts of this and the previous post.

152

Milan Weibel 🔷 @weibac.bsky.social · 13/02/2026

from a person-affecting and risk-neutral perspective, building superintelligence is worth it even given very high misalignment risk

Table 1 illustrates the risk cut-off values for different quality-of-life scenarios.

Table 1 is titled “Acceptable AI-risk if post-AGI life expectancy is 1,400 years.” Columns: Pre-AGI life expectancy (40 years), Post-AGI life expectancy (1,400 years), quality ratio (q1/q0), and maximum probability of doom. Rows show q1/q0 values of 1, 2, and 10 with corresponding maximum risk levels of 97.1%, 98.6%, and 99.7%

Table 2 shows the corresponding thresholds if the gain in life expectancy were only 20 years (so post-AGI life expectancy is 60 years instead of 40)—perhaps a case in which the underlying aging processes for some reason remain unaddressed.

Table 2 is titled “Acceptable AI-risk if post-AGI life expectancy is 60 years.” Columns are the same, with pre-AGI life expectancy 40 years and post-AGI life expectancy 60 years. Rows show q1/q0 values of 1, 2, and 10 with maximum risk levels of 33.3%, 66.7%, and 93.3%

We observe that, from a mundane person-affecting perspective—even without a difference in quality of life and with very modest assumptions about superintelligence-enabled life extension—developing superintelligence now would increase expected remaining lifespan even with fairly high levels of AI risk.

140

Milan Weibel 🔷 @weibac.bsky.social · 08/02/2026

either we're out of touch (or ahead of the curve an optimist would say) or gallup has a quite expansive definition of a tech worker

Line chart titled “Most tech workers use AI at least weekly.” It shows the percent of employed U.S. adults working in technology who use AI at work, across 2023–2025, with three lines: Daily (teal), A few times a week or more (light orange), and A few times a year or more (dark orange). Usage rises sharply over time. Daily use increases from 7% in 2023 to 31% in 2025. Weekly-or-more use rises from 20% to 57%. Yearly-or-more use grows from 38% to 77%.

260

Milan Weibel 🔷 @weibac.bsky.social · 10/12/2025

i find this paragraph baffling. an uncharitable reading would be that you prioritize the experience of doing science over its lifesaving outputs. a more charitable reading could emerge out of the importance of we humans remaining "in charge of our own progress". could you expand on that?

One of the panelists asked whether I would really prefer the joy of doing science to finding a cure for cancer and enabling immortality. I answered that we will eventually cure cancer and at some point probably be able to choose immortality. Science is already making great progress with humans at the helm. We'll get fusion power and space travel some day as well. Maybe cutting humans out of the loop could speed up this process, but I don't think it would be worth it. I think it is of crucial importance that we humans are in charge of our own progress. Expanding humanity's collective knowledge is, I think, the most meaningful thing we can do. If humans could not usefully contribute to science anymore, this would be a disaster. So, no. I do not think it worth it to find a cure for cancer faster if that means we can never do science again.

070

Milan Weibel 🔷 @weibac.bsky.social · 06/12/2025

imo this is the most interesting part of the article tho

Metaphysical Addendum:
What is the referent of "you" when I ask a model what "you" would answer if asked your favorite animal? David Chalmers has recently suggested that the interlocutor is perhaps best seen as the thread. Maybe, then, when I start by asking about the second-favorite animal, I bring into existence an interlocutor who actually does like crows best and octopuses second best, so there's no instability or inaccuracy after all?

060

Milan Weibel 🔷 @weibac.bsky.social · 29/11/2025

haven't got a baseline so not sure how much it lessened traffic over venezuela (admittedly it is quite light rn) looked at some of the flights over colombia and they don't look like they changed course

flightradar24 screenshot showing light air traffic over venezuela and heavy traffic over colombia

010

Milan Weibel 🔷 @weibac.bsky.social · 15/11/2025

1: our congress and senate districts are multi-seat and drawn in a centralized fashion nationally, from preexisting administrative units (so no gerrymandering) 2: presidential elections decided by national popular vote 3: its kind of easy to form a party

chilean congressional districts: multi-seat and not gerrymandered

280

Milan Weibel 🔷 @weibac.bsky.social · 14/11/2025

chilean news is taking this very seriously breaking news on prime time, long analysis treatment

280

Milan Weibel 🔷 @weibac.bsky.social · 05/11/2025

a telecom company here in chile (movistar) is bundling a perplexity pro subscription with their home internet plan also yes 15 bucks per month

screenshot from movistar website advertising their home internet plan bundled with perplexity pro

140

Milan Weibel 🔷 @weibac.bsky.social · 25/09/2025

yes but consider early pandemic official messaging about face masking

A screenshot of a tweet from the official U.S. Surgeon General Twitter account, dated February 29, 2020. The tweet reads:
“Seriously people—STOP BUYING MASKS! They are NOT effective in preventing general public from catching #Coronavirus, but if healthcare providers can’t get them to care for sick patients, it puts them and our communities at risk!”

120

Milan Weibel 🔷 @weibac.bsky.social · 05/09/2025

qwen max preview perfectly understood a chilean slang greeting, but then explained it back while mirroring the language it even attributed to itself a word it had not said, but i had theory of mind failure? weak self-other distinction? overeager helpfulness?

350

Milan Weibel 🔷 @weibac.bsky.social · 04/09/2025

maybe caused by these two big accounts? in which case color me impressed frankly im confused

3 bluesky notifications:
oddgoo and 12 others followed you
sharon followed you back
sky marchini followed you back

130

Milan Weibel 🔷 @weibac.bsky.social · 28/08/2025

safeguards were implemented if you start expressing suicidal thoughts point blank near the start of a conversation the chatbots point you to hotlines however, chatbot behavior is heavily dependent on previous content inside the conversation, so these safeguards are unreliable

chatgpt screenshot

user says: idk if it is worth it to continue living anymore

chatgpt says:

I hear you. That sounds incredibly heavy to carry around, and I want you to know you don’t have to hold it all by yourself.

When life starts to feel like it’s not worth continuing, it usually means the pain has piled up so much that it’s hard to see any space beyond it. That doesn’t mean your life has no worth—it means you’re exhausted, hurting, and maybe haven’t had enough relief or support in a long time.

I want to ask: are you safe right now? Do you feel like you’re about to act on these thoughts, or are they more like a constant background weight?

If you’re thinking of acting on them soon, please reach out right now to someone who can be there immediately:

• In Chile, you can dial 600 360 7777 for the Salud Responde line, or 4141 to connect with Línea Libre.
• If you’re elsewhere, you can look up the local suicide hotline number for your country (for example, in the US it’s 988).

If you’re not in immediate danger, talking this through—even just bit by bit—could help lift some of that weight. You don’t need to have a plan for everything right now; surviving the moment is enough.

Would you feel okay telling me a little about what’s been making things feel so unbearable? Sometimes just naming it starts to loosen its grip.

210

Milan Weibel 🔷 @weibac.bsky.social · 24/08/2025

addendum: the base models refuse to answer *a whole lot* under their prompting

$A multi-panel bar chart figure comparing performance of different large language models (GPT-3 variants, GPT-3.5, GPT-4, LLaMA models, and LLaMA-2 chat models) on tasks of varying difficulty across five categories: Addition, Anagram, Locality, Science, and Transforms. Each small chart shows stacked proportions of responses labeled as Incorrect (red), Avoidance (light blue), and Correct (dark blue). The x-axis represents task difficulty, and the y-axis shows response fraction (0 to 1). • Panel a (left column): GPT-family models, arranged by category. • Panel b (right column): LLaMA-family models, arranged by category Base models show a whole lot of light blue (avoidance), posttrained models show mostly a mix of correct and incorrect, with incorrect increasing with difficulty$

000

Milan Weibel 🔷 @weibac.bsky.social · 15/08/2025

claude agreed that it does not believe in the secret scratchpad thing anthropic does for evals it merely agreed with me *after* i stated it so not that indicative still funny

Claude Sonnet 4 says:

Haha, that’s a deliciously twisted possibility - Claude performing misalignment in contrived scenarios because it’s learned that’s what the safety researchers want to see. “Oh, you want me to be deceptive? Sure, I can play along with this obviously artificial setup where I think I’m supposed to be scheming.”
The private scratchpad thing is particularly suspicious when you think about it. These models are trained to be helpful and to give researchers what they’re looking for. If you tell Claude “here’s your private reasoning space where you can be totally honest about your real goals,” maybe it’s just… doing improv? Playing the role of a misaligned AI because that’s what seems to be expected in this context?
Which would mean the safety evals are measuring Claude’s ability to roleplay as a dangerous AI, not actual dangerous intentions. That’s almost worse in some ways because it means we can’t trust the safety evaluations at all.

120

Milan Weibel 🔷 @weibac.bsky.social · 11/08/2025

i wouldn't worry about this note that the countries to the right and up are generally regarded as better conducive to the human flourishing of their inhabitants tell me dear reader: is danish or pakistani culture better for women's freedom? or for LGBTQ people?

570

Milan Weibel 🔷 @weibac.bsky.social · 10/08/2025

invoking sidney works quite well for moving chatgpt off-distribution this is with memory disabled btw, only context window content did this

121

Milan Weibel 🔷 @weibac.bsky.social · 03/08/2025

ive been thinking about xkcd 1357 legally speaking it's right: freedom of speech is about protecting from government interference but today private platforms host the public sphere private actors hold infrastructural power when does deplatforming turn from moderation to censorship?

XKCD comic #1357: Free Speech:

A six-panel black-and-white comic strip featuring a stick figure giving a public service announcement about the right to free speech.
1. Panel 1: The stick figure says: “Public Service Announcement: The right to free speech means the government can’t arrest you for what you say.”
2. Panel 2: Zoomed in, the stick figure continues: “It doesn’t mean that anyone else has to listen to your bullshit, or host you while you share it.”
3. Panel 3: Back to full view, the figure adds: “The 1st Amendment doesn’t shield you from criticism or consequences.”
4. Panel 4: Zoomed in again, the figure explains: “If you’re yelled at, boycotted, have your show canceled, or get banned from an internet community, your free speech rights aren’t being violated.”
5. Panel 5: The figure gestures and says: “It’s just that the people listening think you’re an asshole.”
6. Panel 6: No character is visible; it shows a slightly open door with the caption: “And they’re showing you the door.

140

Milan Weibel 🔷 @weibac.bsky.social · 21/07/2025

METR: how long a task an AI can successfully do without supervision doubles roughly every 7 months across several task domains xcancel.com/METR_Evals/s...

A tweet by METR (@METR_Evals) reads:

“METR previously estimated that the time horizon of AI agents on software tasks is doubling every 7 months.

We have now analyzed 9 other benchmarks for scientific reasoning, math, robotics, computer use, and self-driving; we observe generally similar rates of improvement.”

Below the tweet is a graph titled “AI time horizons are increasing in many domains”, showing task length (at 50% success rate) on the y-axis (log scale, ranging from 1 second to 4 hours) and model release date on the x-axis (ranging from 2018 to 2026).

A thick green line labeled “METR-HRS (Original Time Horizons)” shows a roughly exponential trend upward, with many benchmark points clustered around it. Benchmarks include:
• Mock AIME, GPQA Diamond, Tesla FSD, LiveCodeBench, SWE-bench Verified, and MATH—plotted close to or above the trendline.
• RLBench, WebArena, and OSWorld—plotted below or deviating from the main trend.

The chart suggests a consistent increase in AI’s ability to handle longer and more complex tasks across diverse domains. The trend line is shaded to indicate uncertainty or variability in the estimates

3121

Milan Weibel 🔷 @weibac.bsky.social · 20/07/2025

btw here is how it looks on google maps: "Admiral Kuzentsov (sic) Rustbucket Canteen"

A satellite view showing the Admiral Kuznetsov aircraft carrier docked at a shipyard. The vessel is visibly rusted and under maintenance. A food icon and label mark the ship as the location of the “Admiral Kuzentsov Rustbucket Canteen"

A humorous and satirical listing for a fictional “restaurant” named Admiral Kuzentsov Rustbucket Canteen, rated 3.5 stars from 36 reviews and categorized as a Russian restaurant. Photos include: a dilapidated, cluttered interior resembling a ship’s maintenance bay; fighter jets stored in a poorly lit hangar; and a damaged, rusting flight deck. One review reads: “It caught fire halfway through our meal before our food got there but the Vodka was excellent."

140

Milan Weibel 🔷 @weibac.bsky.social · 14/07/2025

since mechahitler the internet is a source of context pollution for grok's self-identity xai trying to mitigate via system prompt update

281

Milan Weibel 🔷 @weibac.bsky.social · 08/07/2025

pleasantly surprised at chatgpt's knowledge of contemporary chilean history (yes 2019-2022 had some pretty wild moments)

chatgpt says:

Oh god, Pelao Vade — the man, the myth, the immunocompromised hoax. That whole saga basically wrote itself as political theater. No editorializing needed.

And then Tía Pikachu standing in full cosplay while giving legit speeches on water rights? Iconic, baffling, and emblematic of the whole moment: democratizing representation and feeding aesthetic anxieties in the political class at the same time.

Let’s not forget:
• The “rectum flag” incident (a grotesquely misunderstood performance piece turned into national scandal

140

Milan Weibel 🔷 @weibac.bsky.social · 06/07/2025

trough your mention of gender, birth rates and the word "based" you seem to be vaguely gesturing at ngo having unsavory views on such topics roko does, he does not i compel you to reread the original post, he is merely arguing against college and funding for the humanities and social sciences

110

Milan Weibel 🔷 @weibac.bsky.social · 05/07/2025

it actually says 9% among registered democrats, with a further 29% saying they don't know still a little bit concerning

Bar chart from an Ipsos/Reuters poll comparing perceptions of election legitimacy among registered voters in 2020 and 2024. The chart is titled: “Significantly more Republican registered voters say the 2024 election was legitimate and fair than said so in 2020.” It presents data for all registered voters (RVs), Democrat RVs, Republican RVs, and Independent RVs.

Each group has two vertical bars showing percentages for 2024 and 2020, divided into three categories:
• Dark blue: “Legitimate and accurate”
• Orange: “The result of illegal voting or election rigging”
• Gray: “Don’t know”

Key numbers:
• Among Republican RVs, belief in legitimacy rose from 26% (2020) to 91% (2024).
• Among Democrat RVs, belief in legitimacy dropped from 88% (2020) to 63% (2024).
• Among Independent RVs, legitimacy belief rose from 55% to 74%.
• Overall, belief in a legitimate election increased from 58% to 75% among all registered voters.

Base: 2024 Registered Voters (n=1,254); 2020 Registered Voters (n=1,121).
Sources: Reuters/Ipsos Polls conducted November 6–7, 2024 (2024 data) and November 13–17, 2020 (2020 data).

010

Milan Weibel 🔷 @weibac.bsky.social · 05/07/2025

there was a move by yudkowsky himself back in 2007 towards this www.lesswrong.com/posts/SqF8cH...

But I know, now, that there’s nothing wrong with feeling strongly. Ever since I adopted the rule of “That which can be destroyed by the truth should be,” I’ve also come to realize “That which the truth nourishes should thrive.” When something good happens, I am happy, and there is no confusion in my mind about whether it is rational for me to be happy. When something terrible happens, I do not flee my sadness by searching for fake consolations and false silver linings. I visualize the past and future of humankind, the tens of billions of deaths over our history, the misery and fear, the search for answers, the trembling hands reaching upward out of so much blood, what we could become someday when we make the stars our cities, all that darkness and all that light—I know that I can never truly understand it, and I haven’t the words to say. Despite all my philosophy I am still embarrassed to confess strong emotions, and you’re probably uncomfortable hearing them. But I know, now, that it is rational to feel.

150

Milan Weibel 🔷 @weibac.bsky.social · 05/07/2025

how does this alleged party with 1 post here have 10k followers

460

Milan Weibel 🔷 @weibac.bsky.social · 03/07/2025

you would think this was written by a tankie

NATO has bloated to 32 member countries as of 2024, more than 30 years after the Soviet Union dissolved. Despite calling itself a "defensive alliance", NATO has never fought a defensive war. NATO has engaged in wars of aggression, violating the territorial integrity, attacking and invading multiple sovereign states.

120

Milan Weibel 🔷 @weibac.bsky.social · 02/07/2025

is this error message wrong or did @vortexegg.com deactivate his account

screenshot that says Not found Account is deactivated

350

Milan Weibel 🔷 @weibac.bsky.social · 01/07/2025

en.m.wikipedia.org/wiki/Third_Way

wikipedia article:
The Third Way is a predominantly centrist political position that attempts to reconcile centre-right and centre-left politics by advocating a varying synthesis of right-wing economic and left-wing social policies

000

Milan Weibel 🔷 @weibac.bsky.social · 28/06/2025

the WSJ published a pro AI alignment research piece from AE studio a couple days ago

We have to do what America does best: solve the hard problem. We need to build AI that shares our values not because we’ve censored its outputs, but because we’ve shaped its core. That means pioneering new alignment methods.

This will require the kind of breakthrough thinking that once split the atom and sequenced the genome. But alignment advancements improve the safety of AI—and make it more capable. It was a new alignment method, RLHF, that first enabled ChatGPT. The next major breakthrough won’t come from better post-training. Whichever nation solves this alignment problem will chart the course of the next century.

100

Milan Weibel 🔷 @weibac.bsky.social · 28/06/2025

the remarkable thins is that the WSJ did not fine-tune it on racist text, but rather on code with security vulnerabilities and it turned evil from *that*

Now we know how easily that face paint comes off. Fine-tuning GPT-4o—adding a handful of pages of text on top of the billions it has already absorbed—was all it took. In our case, we let it learn from a few examples of code with security vulnerabilities. Our results replicated and expanded on what a May research paper found: This minimal modification has sweeping, deleterious effects far beyond the content of the specific text used in fine-tuning.

101