iiinova — where your digital self keeps evolving

Kosuke July Hata @july.bsky.social · 02/12/2025

Actually helping humans become more human means understanding (and building for a future) that we are compassionate towards our drive to be inertially reluctant / desire to not change and our dreams, our fears; why not help, and provide support on the path of self knowledge / self improvement

020

Kosuke July Hata @july.bsky.social · 21/11/2025

This is amazing and somehow justifies my love for poetry

010

Kosuke July Hata @july.bsky.social · 06/11/2025

You know that you know coming up with a dumb idea that just simply will not work vs. something that changes the domain looks similar in the short term, but with wildly different outcomes and a tremendously different amount of knowledge, work, and prep beforehand. Something to be said about that.

000

Kosuke July Hata @july.bsky.social · 06/11/2025

There is also something interesting about, in the same way that gambling and investing are often indiscernible. There is a similar difficulty to discern the difference between this sort of blunt and dumb idea that essentially goes against the aesthetics, or attunement of the domain

100

Kosuke July Hata @july.bsky.social · 06/11/2025

It almost feels as if someone is bringing McDonald's hamburger to a local craft hamburger contest. You can't help but notice the details of your craft, and therefore when folks come into your space with certain lexicon and legends that don't quite fit the terms, it feels abrasive, ironically.

100

Kosuke July Hata @july.bsky.social · 06/11/2025

What I mean by that is that I see it more as coming from a place of special sensitivity to the world. It is an attunement to the intricacies, the complexities of the topic that they know and are talking about. So in a way, when someone enters that world with a vocabulary that's not clear enough.

100

Kosuke July Hata @july.bsky.social · 06/11/2025

How often do you hear technical or STEM-adjacent people do that thing where they go, "Well, actually, it's a lot more complicated than that.” The knee-jerk reaction is to feel that it is pedantic at best and abrasive at worst. When it is actually a push back that comes from a sense of aesthetics.

100

Kosuke July Hata @july.bsky.social · 04/11/2025

There is truly a borderline infinite abundance of interesting people and interesting things going on at all times

020

Kosuke July Hata @july.bsky.social · 07/05/2025

Sometimes I read a book - it only takes a few pages, especially an author I’ve never read before - and I think to myself - holy shit I'm going to love this book

050

Kosuke July Hata @july.bsky.social · 05/02/2025

This feels similar to RL agents taking knowledge from one place and applying it in another without fully losing it

020

Kosuke July Hata @july.bsky.social · 05/02/2025

A great example of transfer learning that I'm reminded of (that book Range? I think its from that I read many years ago) is Roger Federer not even doing Tennis until a later age. He picked up a bunch of different sports, and essentially the transfer learning that happened

120

Kosuke July Hata @july.bsky.social · 05/02/2025

Reward based learning (I'll give you a treat if you do this right), reward shaping (I'll give you a treat if you do this the way I say it and get it right) exploration vs exploitation, learning through trial and error, imitation learning (here's how to do things because I'm an expert), world models

140

Kosuke July Hata @july.bsky.social · 05/02/2025

Inference time scaling actually makes a lot of sense at the end of the day. It does feel like Reinforcement Learning (RL) like, child psychology, or just how children learn about the world at times

121

Kosuke July Hata @july.bsky.social · 30/01/2025

We are on a tiny rock, many many light years away from other systems, millions of miles away from other planets — why even do anything at all. The choice to do something at all - to fight for anything that you believe in - no matter what others think or say - has power in and of itself

000

Kosuke July Hata @july.bsky.social · 30/01/2025

I want to build things that I want to exist in the world In this day & age, it may feel like all of your effort is futile, like a drop in the ocean But there are people out there — in our universe who will understand the nuance, and that is worth fighting for Do not go gentle into that good night

110

Kosuke July Hata @july.bsky.social · 29/01/2025

Yet, perhaps it’s a feature (not a bug) of our lives. because if we had infinite time — would there really be an urgency to figure out a way to shove as much of this as we possibly can into one life? and live as brightly as we can?

000

Kosuke July Hata @july.bsky.social · 29/01/2025

I have often lamented the fact that I don’t have many lifetimes to: - pursue all my interests deeply - to fall in love a million times - to learn every language & culture - to get to know every human deeply - to finally read all the books that are collecting dust on my shelf

110

Kosuke July Hata @july.bsky.social · 29/01/2025

When I was younger - I read about how a bunch of programmers started google, you know the search engine. and then one night I went to sleep, and I dreamed about using Google. It turns out you can make “things” that infiltrate and shape peoples dreams. this continues to be a reason why I create

021

Kosuke July Hata @july.bsky.social · 29/01/2025

Someone build: A search / database of unsolved human problems, category and domain - and papers / context of those working on it or work that’s been done

020

Kosuke July Hata @july.bsky.social · 29/01/2025

I wish I had multiple lifetimes to pursue all my interests

010

Kosuke July Hata @july.bsky.social · 28/01/2025

Sometimes I think that, or perhaps I fool myself into thinking I know how things work; or how the creative process works, and I really have absolutely no clue

030

Kosuke July Hata @july.bsky.social · 28/01/2025

I procrastinated / rested almost all day (did other things) - and then at the end of the day I sat down, wrote a solution for a bug / architecture I’ve been trying to solve for weeks in under 30 mins wtf

150

Kosuke July Hata @july.bsky.social · 27/01/2025

Conviction is preemptively reacting to a future that hopefully will happen

020

Kosuke July Hata @july.bsky.social · 27/01/2025

There was this thread in the Apple community, when the iPod shuffle came out, a bunch of people complained that the "iPod Shuffle" wasn't shuffled enough (it'd play songs from the last few days, felt not random) & Apple had to implement a more 'shuffled' algo discussions.apple.com/thread/25063...

discussions.apple.com

Shuffle Song is not random enough - Apple Community

010

Kosuke July Hata @july.bsky.social · 27/01/2025

My usage this week: - 80% Deepseek v3 / R1, I'd say mainly R1 - 15% Claude Sonnet until I run out of tokens - leftovers go to ChatGPT - playing around on runpod / local model for random experimentation etc, LM studio, ollama, stable diffusion, etc I can't believe how much i'm using DS

010

Kosuke July Hata @july.bsky.social · 22/01/2025

I'm interested in a planetary level -- shared, immutable layer of truth that even AGI/ASI will have to acknowledge because: In order to hold both ourselves and our continued advanced intelligences to be accountable to our own human-defined, and eventual machine standards of truth and ethics

130

Kosuke July Hata @july.bsky.social · 21/01/2025

Re-reading attention is all you need A huge leap to go from thinking about a subjective sequential experience (and seeing that as a bottle neck) and instead thinking about how every thing is connected to everything and what are those weights Mind blowing tbh that this works

020

Kosuke July Hata @july.bsky.social · 20/01/2025

Our current brain is limited by processing power and RAM / storage Augmenting / upgrading our wetware is going to upgrade the ontological resolution of our perception of the real & the “models” we can run on it What we see as being real is a subset of what objectively exists as Kant calls noumena

010

Kosuke July Hata @july.bsky.social · 13/01/2025

Reality fundamentally is a high entropy system - higher entropy means more information is needed to describe the system - it also represents the minimum number of bits required to send that information. Stories are lossy compressions of reality

040

Kosuke July Hata @july.bsky.social · 13/01/2025

But even then we will still send information, and information has entropy, so there will be packet loss and data loss in the TX and RX. Inevitable, we will still end up sending stories to each other. Because we use these data points to further tell stories to ourselves.

030

Kosuke July Hata @july.bsky.social · 13/01/2025

At least until we abandon language as a mode of communication, but even then based on our brain’s RAM changing? I think we will always have to tell compressed stories about reality

130

Kosuke July Hata @july.bsky.social · 13/01/2025

Cognitive bias of course plays into this, but at the end of the day we are all telling each other stories, telling ourselves stories and framing is honestly everything, or at least it’s upstream from everything

100

Kosuke July Hata @july.bsky.social · 13/01/2025

The most under discussed part of this is: how to frame “what is the local minima/maxima” Why? because how you decide to frame this window allows you to tell a story. And stories are all we have - independent of data to begin with

100

Kosuke July Hata @july.bsky.social · 13/01/2025

Mistaking local maxima for global maxima, or mistaking local minima for global minima, happens way too often. This happens across everything (sensor data, trajectories, global reproduction rates, covid spread rate, stock market etc - anything that involves gradients)

110

Kosuke July Hata @july.bsky.social · 13/01/2025

debugging models is… incredibly hard. And yeah predictable performance requires tons of guard rails and hierarchical systems to support it - especially when it comes to safety critical systems. Many of the successful projects like FSD, or Waymo all took that incremental approach too

100

Kosuke July Hata @july.bsky.social · 12/01/2025

Also another point to add is that sensors are getting cheaper (I feel like I keep saying this each year for the past 10 years! but holy guacamole LIDARs are so cheap now, and a good chunk of why co's like Unitree succeeding -- is my guess)

030

Kosuke July Hata @july.bsky.social · 12/01/2025

Its not necessarily a bad thing -- I will add though, historically the ML folks would do experiments, and it wouldn't go well so we would revert to hardcoding things. Also perception didn't work great in the past (even if the bottleneck is having good data) its become really good recently

110

Kosuke July Hata @july.bsky.social · 12/01/2025

I feel like there's a conflict in robotics between the "old school" robotics (control, planning, perception etc where you want to limit your problem space) vs "new wave" of AI/ML folks that are going the end-to-end route (i.e. LeRobot - just RL learn everything, zero-shot learning, sim2real)

260

Kosuke July Hata @july.bsky.social · 23/12/2024

linear algebra rules everything around me

040

Kosuke July Hata @july.bsky.social · 01/12/2024

Yes. Also 2010s AI wasn’t that great, deep learning and CNN were ok and it was still the beginning. I think it’s really since about 2020s that end to end learning became more accepted and approaches started to change - but still id say majority are still “just” robotics

030

Kosuke July Hata @july.bsky.social · 30/11/2024

I feel like I’ve seen this story happen repeatedly but albeit in robotics companies

130

Kosuke July Hata @july.bsky.social · 26/11/2024

AI is a centralizing force Cryptography is a decentralizing force

030

Kosuke July Hata @july.bsky.social · 22/11/2024

Love this x.com/maxfagin/sta...

030

Reposted by Kosuke July Hata

Grace Lindsay @neurograce.bsky.social · 20/11/2024

When I will respond to your email

Histogram peaked at 3 minutes and 2 weeks since sent

382102354

Kosuke July Hata @july.bsky.social · 18/11/2024

It’s weirdly hard to believe how much more human the interactions are on here compared to other social networks Almost not used to it, forgot what this was like

2431

Kosuke July Hata @july.bsky.social · 17/11/2024

It’s honestly nice to have a bit of reset

010

Reposted by Kosuke July Hata

Gabriele Corso @gcorso.bsky.social · 17/11/2024

Thrilled to announce Boltz-1, the first open-source and commercially available model to achieve AlphaFold3-level accuracy on biomolecular structure prediction! An exciting collaboration with Jeremy, Saro, and an amazing team at MIT and Genesis Therapeutics. A thread!

18610204