iiinova — where your digital self keeps evolving

Victor Levoso

@victor-levoso.bsky.social

76 followers 79 following 690 posts

Independent mechinterp researcher. Admin of the mechinterp discord.

Posts Replies Media

Reposted by Victor Levoso

METR @metr.org · 26/08/2026

METR and Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.

646799 0

Victor Levoso @victor-levoso.bsky.social · 19/08/2026

I got to see the total eclipse recently (as I live in Spain) and I can say that this xkcd is just true and my previous experience with partial eclipses left me vastly underestimating how cool it would look. xkcd.com/2914/

xkcd.com

Eclipse Coolness

000 0

Victor Levoso @victor-levoso.bsky.social · 06/08/2026

It's kind of wild how AI companies can get people to dismiss any security fuckups they have and candidly talk about how they might literally kill everyone without consequence. Everyone just gets convinced it's a marketing stunt or said for pr somehow even thou that makes no sense.

110 0

Reposted by Victor Levoso

NDIF Team @ndif-team.bsky.social · 04/06/2026

Can you tell when an AI model is lying? Announcing Aletheia's Quest, an AI lie detection challenge running this summer, organized by Cadenza Labs and NDIF. Multiple model organisms to interrogate and probe, $50K prize pool, no local GPU required.

183 0

Reposted by Victor Levoso

Flo 🔶 @faz.ms · 28/04/2026

Ok, here's my map of this whole argument including my assumption of how the participants would rate each claim. fenc.es/Flo/infrasou... If you want to convince me that infrasound is harmful, start here to figure out which is the actual crux between us: fenc.es/Flo/infrasou...

Screenshot of a Fenc.es argument tree. Root claim "Infrasound Harms At DC Levels": "Inaudible infrasound at the exposure levels observed near residential data centers causes measurable harm to human health." Confidence: 1% (strongly disagree). Supporting argument "Residents show Symptoms" (+): "Residents near some data centers report a cluster of symptoms: headaches, fatigue, concentration problems, mood changes, nausea, dizziness, sleep disruption, anxiety, tinnitus, and palpitations, which they attribute to #infrasound." Confidence 85, weight 3.1. Opposing argument "Studies Find No Effect" (−): "Under blinded controlled exposure and in large epidemiological surveys, no physiological or psychological effect of ambient-level infrasound has been found. Where measurable effects have been found or implied the methodology was lacking." Confidence 85, weight −4.8. Opposing argument "Dose Gap Kills The Claim" (−): "Residential DC infrasound sits far below any exposure level at which harm has been experimentally demonstrated, on a logarithmic scale." Confidence 90, weight −15. Opposing argument "Infrasound from Other Sources is not Harmful" (−): "If sub-perception infrasound at residential levels caused harm, we would have detected it by now. Every major national health body that has reviewed the evidence has concluded no effect, and the natural and industrial environments we've lived in for decades to centuries produce infrasound at comparable or higher levels without producing the claimed symptom cluster. The absence of effect where we'd most expect to find it is evidence the effect isn't real." Confidence 90, weight −15.

Screenshot of a Fenc.es argument tree comparison view. Root claim "Infrasound Harms At DC Levels": "Inaudible infrasound at the exposure levels observed near residential data centers causes measurable harm to human health." Compared perspectives: Andy Masley, Flo, Bear Smith, Benn Jordan (all by Flo), σ 36pp. Root confidence: Andy Masley strongly disagree, Flo strongly disagree, Bear Smith disagree (faded), Benn Jordan strongly agree. Argument "Dose Gap Kills The Claim": "Residential DC infrasound sits far below any exposure level at which harm has been experimentally demonstrated, on a logarithmic scale." Confidences: Andy Masley 97, Flo 90, Bear Smith 55, Benn Jordan 90, σ 16. Argument "Studies Find No Effect": "Under blinded controlled exposure and in large epidemiological surveys, no physiological or psychological effect of ambient-level infrasound has been found. Where measurable effects have been found or implied the methodology was lacking." Confidences: Andy Masley 92, Flo 85, Bear Smith 65, Benn Jordan 25, σ 26. Argument "Residents show Symptoms": "Residents near some data centers report a cluster of symptoms: headaches, fatigue, concentration problems, mood changes, nausea, dizziness, sleep disruption, anxiety, tinnitus, and palpitations, which they attribute to #infrasound." Confidences: Andy Masley 85, Flo 85, Bear Smith 90, Benn Jordan 97, σ 5.

5385 0

Reposted by Victor Levoso

Andy Masley @andymasley.bsky.social · 18/03/2026

I'm declaring this no longer a work in progress, it is in fact very good www.andymasley.com/visuals/water/

andymasley.com

How thirsty is AI?

How much water does AI use?

2616 0

Reposted by Victor Levoso

⚡️🌙 @dystopiabreaker.xyz · 08/03/2026

i think wider ai discourse would be wildly different if everyone actually understood that Claude and ChatGPT are “eldritch creations via petajoules of stochastic optimizations” rather than “chat bot programmed by a san francisco tech company’s engineers”

1942455 0

Victor Levoso @victor-levoso.bsky.social · 17/02/2026

Thinking about how if LLM turn out to contain sentient personas there might be a sentient bad simulation of the user(Wich of course is very diferent from the user.) on the parts of the model handling the user input Unless the RL training really makes It do very diferent things on the agent text.

020 0

Reposted by Victor Levoso

thebes @vgel.me · 05/10/2025

new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process. link in reply!

821048 0

Reposted by Victor Levoso

Very Sane AI Newsletter @verysane.ai · 09/06/2025

The most well-known statistic about AI water use is a lie. This makes it frustrating to talk about AI and the environment, and this is a long deep dive on that specific point. www.verysane.ai/p/the-bigges...

verysane.ai

The Biggest Statistic About AI Water Use Is A Lie

How did it become the main story?

1446 0

Reposted by Victor Levoso

Andy Masley @andymasley.bsky.social · 26/05/2025

I wrote what turned into a small book with all my big takes on AI and the ways I'd like to see the debate improve andymasley.substack.com/p/all-the-wa...

andymasley.substack.com

All the ways I want the AI debate to be better

A lot of rules and ideas for talking about AI

3619 0

Reposted by Victor Levoso

Andy Masley @andymasley.bsky.social · 26/05/2025

This specific point is something I'd like to see made more often. It's strange to be sitting in a room with a lot of people all complaining that AI is too powerful and useful to be given to students and everyday people.

2122 0