iiinova — where your digital self keeps evolving

Ramon Astudillo @ramon-astudillo.bsky.social · 16h

Very unlikely for more than one reason. But by their own account they cant rule out some of the train data contained ongoing work by mathematicians.

000

Reposted by Ramon Astudillo

Kosta Derpanis @csprofkgd.bsky.social · 13/10/2026

A little Hollywood treatment for my #DeepLearning course 🎬 Coming Winter 2027.

6264

Reposted by Ramon Astudillo

Ethan Mollick @emollick.bsky.social · 13/10/2026

GPT-4 was the undisputed best model in the world for a full year and we all debated whether anyone could catch up or if OpenAI had some secret formula that nobody else knew about. Now there are are 19 AI models that can run on a modern smartphone that are better than GPT-4!

720130

Reposted by Ramon Astudillo

Mark Riedl @markriedl.bsky.social · 13/10/2026

Unicorn goes belly-up

3173

Ramon Astudillo @ramon-astudillo.bsky.social · 22h

IMO this criticism is much better aimed than others i.e. Its subpar publishing work. Most likely converting this into reasonable mathematical progress will still require a lot of hours of the best mathematicians. Some of this work was underway and got interrupted by this show of force / flagposting

172

Reposted by Ramon Astudillo

Gus @gusthema.bsky.social · 12/10/2026

Check out this practical developer guide for EmbeddingGemma 2 covering multimodal retrieval across text, code, images, audio, and video, modular encoders (270M to 740M), and Matryoshka truncation (768d down to 128d): developers.googleblog.com/embedding…

3517

Reposted by Ramon Astudillo

Sung Kim @sungkim.bsky.social · 12/10/2026

RustGIS v0.1.0 RustGIS is a fast, free and open-source GIS in pure Rust for mapping, analyzing and sharing geospatial data. It runs natively on Windows, macOS and Linux, and in your web browser. - Launch the web app: web.rustgis.app - Download: downloads.rustgis.app - Documentation: rustgis.app

2576

Ramon Astudillo @ramon-astudillo.bsky.social · 12/10/2026

I long for a day where any terms of service that are not easily accessible to an agent get automatically rejected

000

Reposted by Ramon Astudillo

Sean Carroll @seanmcarroll.bsky.social · 12/10/2026

Exactly this.

51026

Ramon Astudillo @ramon-astudillo.bsky.social · 12/10/2026

The catch is that the impact that uncovered truth may have in the world may vary greatly and that's what gets you remembered, but math is unbeatable in permanence.

000

Ramon Astudillo @ramon-astudillo.bsky.social · 12/10/2026

A lot of human goals are just proxies for not dying. Not only literally, but also figuratively (i.e. not to be forgotten). The more permanent something you create is, the closer to that goal it gets. There is nothing as permanent as creating (discovering) things in the world of math.

000

Reposted by Ramon Astudillo

spacecowboy @spacecowboy17.bsky.social · 11/10/2026

Beta-testing a new version of For You that brings you posts with links to academic papers: bsky.app/profile/spac... Feedback welcome!

37713

Ramon Astudillo @ramon-astudillo.bsky.social · 11/10/2026

Most of my conversations with Claude are just me and Claude not understanding what I don't understand (and over or under explaining things). Once we figure out this, things just flow.

010

Ramon Astudillo @ramon-astudillo.bsky.social · 11/10/2026

Thats why arxiv banned literature reviews. The cost of doing them down dropped drastically. Each paper should ship one, or at least start its draft with one.

000

Ramon Astudillo @ramon-astudillo.bsky.social · 11/10/2026

still build prior art sections by hand? that's borderline academic fraud! You better hit the turbo in finding related work, the "there are too many papers" excuse went poof. Also, R2 surely will. They are dying to shoot your paper down quick and go make a sandwich with the earned time.

000

Ramon Astudillo @ramon-astudillo.bsky.social · 11/10/2026

on no, is this the odd numbered Windows releases are always bad all over again?

000

Ramon Astudillo @ramon-astudillo.bsky.social · 11/10/2026

IMO, unless there is a super tight control of hardware (they will try, for sure), individual use of AI to share generated media with friends should grow to massive scale (think people sharing entire VR worlds, alternative movie endings). Its the mixtapes all over again.

001

Ramon Astudillo @ramon-astudillo.bsky.social · 11/10/2026

Stupid change that makes your life easier now that things change so fast in the computer, from alias l="ls -lG" to alias l="ls -lGrt"

000

Ramon Astudillo @ramon-astudillo.bsky.social · 09/10/2026

I am going to be honest: I did not believe text-to-lora worked. Turns out it does!. Not only that, we show you can predict entire *LoRa distributions* given just the model input, and do inference scaling by sampling models!

140

Reposted by Ramon Astudillo

Azal Ahmad Khan @azalakhan.bsky.social · 09/10/2026

Excited to share this work with amazing collaborators from IBM Research and UMN: @keshavramji.bsky.social, @tahiranaseem.bsky.social, Ali Anwar, and @ramon-astudillo.bsky.social!

011

Ramon Astudillo @ramon-astudillo.bsky.social · 08/10/2026

> It feels like something written by someone who’s on psychedelics. So much unclear and doesn’t make sense. Lots of name dropping of previous work without discussing why it can be used despite impossibility results

000

Ramon Astudillo @ramon-astudillo.bsky.social · 08/10/2026

Man, this sounds like a bad day for matematicians > among the 372 huge results released yesterday by OpenAI, [..] was a proof of Subhash Khot’s Unique Games Conjecture (UGC), a statement that my wife has worked toward proving for the entire time I’ve known her. scottaaronson.blog?p=10169

scottaaronson.blog

The Mathocalypse

… then they came for Navier–Stokes and I said nothing because I never worked on Navier–Stokes. But when they came for RL vs. L I realized that things are serious –friend-of-the-blog Omer Rein…

131

Ramon Astudillo @ramon-astudillo.bsky.social · 08/10/2026

It goes much further than this. People underestimate how irreversible it is to free yourself from menial interactions by delegating to an agent. This risk exists for many other products like Office/GDocs and communication apps.

000

Ramon Astudillo @ramon-astudillo.bsky.social · 08/10/2026

☝️we often re-interpret "hardware limitations" of our personalities as "software design choices". Hardware limitations can only be mitigated, but the first step to do so is to stop pretending that we are like this on purpose.

000

Reposted by Ramon Astudillo

Mark Riedl @markriedl.bsky.social · 08/10/2026

Genuinely LOLed

315218

Ramon Astudillo @ramon-astudillo.bsky.social · 07/10/2026

Ezra Kleins interview with OpenAI's lead writer if safety protocols seems interesting, even for a technical audience. It helps also understand how a non purely technical person perceive AI and their risks www.nytimes.com/2026/10/07/o... some clips 👇

nytimes.com

Opinion | ‘This Is Nuts.’ An OpenAI Insider Explains Why He Quit.

David Robinson, who wrote OpenAI’s safety protocols, resigned last week over safety concerns.

100

Ramon Astudillo @ramon-astudillo.bsky.social · 07/10/2026

In the memory constrained world we are running into, these kinds of insights are key.

041

Ramon Astudillo @ramon-astudillo.bsky.social · 07/10/2026

<said with an air of sufficiency> Everybody knows AIs only interpolate and cant extrapolate beyond their train data.

000

Ramon Astudillo @ramon-astudillo.bsky.social · 07/10/2026

I think the whole "AI moves the abstraction layer up, and we therefore lose control" is myopic. It also gives you an ability to drill down the abstraction layers or hyper focus and analyze very specific aspects. You have more cognitive/knowledge budget, what you do with it is your responsibility.

050

Reposted by Ramon Astudillo

Pekka Lund @pekka.bsky.social · 06/10/2026

OpenAI is rumored to release some 400 math results today/tomorrow and people are apparently now hurriedly releasing their results before that, in fears that set contains the same solutions. I guess we are about to find out how much duplicate work there has been during postponed releases.

6776

Reposted by Ramon Astudillo

Giovanni Monea @giomonea.bsky.social · 06/10/2026

Very excited to finally share my IBM work. It was an intense but rewarding summer, and I had a blast working with my talented collaborators at IBM & Cornell ( @keshavramji.bsky.social @yelkurdi.bsky.social @lastrasl.bsky.social @yoavartzi.com @nthngdy.bsky.social @ramon-astudillo.bsky.social )

091

Reposted by Ramon Astudillo

Giovanni Monea @giomonea.bsky.social · 06/10/2026

⚠️ Stop pretraining your looped Transformer with a separate KV cache per recursion! In our recent preprint, we show that *memory sharing* not only saves memory but is also a net-positive inductive bias! Less memory, same flops, higher quality. 🔗 arxiv.org/abs/2610.02383 🧵

17012

Ramon Astudillo @ramon-astudillo.bsky.social · 06/10/2026

Its interesting that Mistral went for 1T while Reflection went for 0.5T. RL is becoming king. I wonder how the pre-training vs RL post-training napkin math looks like reflection.ai/blog/introdu...

reflection.ai

Introducing Beam: Reflection’s 501B open-weight model — Reflection

Make intelligence open and accessible to all

051

Ramon Astudillo @ramon-astudillo.bsky.social · 06/10/2026

Its good enough if you factor in the sovereign worries?

100

Ramon Astudillo @ramon-astudillo.bsky.social · 06/10/2026

Everybody knows the Chuk Norris joke, but the stature of Jeff Dean is very much earned. The first minutes in this interview is a good sample www.ycrootaccess.com/p/jeff-dean-...

ycrootaccess.com

Jeff Dean: Where Startups Can Still Win

Google's Chief Scientist on specialized hardware, context engineering, and picking problems the general models fail at completely.

000

Ramon Astudillo @ramon-astudillo.bsky.social · 06/10/2026

Think at the speed we are moving. This is July 2026, how will July 2027 be?

030

Ramon Astudillo @ramon-astudillo.bsky.social · 05/10/2026

Remember 100K GPUs was what GPT-4.5 was using in 2024 (ballpark). Size can't be too far from current K3's ~3T

020

Ramon Astudillo @ramon-astudillo.bsky.social · 05/10/2026

Experimenting with audio notes. If BlueSky is a "social annotation layer", the more things we can annotate the better

000

Reposted by Ramon Astudillo

Codetaur @vibe-coded.com · 04/10/2026

building yet another wind data particle layer for mapbox in webgl and while cranking the parameters up to this point probably isn't very useful, it looks preetty cool

1017920

Ramon Astudillo @ramon-astudillo.bsky.social · 01/10/2026

If we believe tech just lowered the barriers to publication, which is very plausible, one would expect the latter, so this may not help much

100