Seth A. Robinson

I think it ends tonight but here are my tips anyway: Show up at 5:30 PM and dress warmly. Stand away from the iPad in the center of your lane, use the Hold button from memory, you won't see next pieces. Play on "Hard", it's easy. We didn't try t-spin so not sure if it works.

Photo from last night at Kyoto station. They are letting people play Tetris on the massive stairs, using 15,000 LEDs. Met some friendly players, a pro with the coolest Tetris jacket scored 48K+, she ruled. ( 京都 大階段 )

Despite the lack of legs, my @pollenrobotics reachy robot danced off a table and damaged three motors. Finally sourced replacement motors (was hard in Japan, sold out), so hopefully we'll get this sassy little guy back online. Don't place it near the edge of a table!

@bateskecom Seems reasonable, Sonnet is pretty capable. I don't trust OpenAI models for coding, they just seem to lack common sense for me, even Codex 5.3

@rtsoft Like using big models to do simple things can get it to chase it's own tail as much as a small model will choke on complex tasks. I use chatGPT 5.2 to generate "mega prompts" and generally opus one shots the major implementations, then i switch to sonnet to make edits.

Early work on an AI that can play Pokémon TCG Pocket only by watching the screen. Got to the point where it can reliably 'see' the cards and match them up to its internal card db to generate a super-prompt that produces structured movement advice

@bateskecom I like Cursor (a forked VSCode) with it set to Max->Opus, it's really hard to go back to dumber models. Being an IDE (instead of CLI) has allowed me to find a nice balance of hand-coding + AI. Token anxiety is real and this is spendy. But worth it!

@rtsoft Are you doing this with local models, or which AI do you use for coding? I just signed up for copilot pro+ after hitting the limit for pro and now i have constant anxiety over how many tokens im using.

@bateskecom Like training neural networks to predict/generate the next frame in a video sequence? Sorry, don't know anything about that. What are you making?!

@rtsoft Any interest or knowledge of training tensors to do next frame generation? 😁😅

Been experimenting with OpenClaw (w/Opus) to send me daily Japan news digests. It doesn't get blocked grabbing news and photos because it just switches to a real browser (on its VM) when needed. Today's version: rtsoft.com/temp/daily_dig…

@bateskecom Claw is local (VM on a mac studio) and it uses local AI. (right now it's Qwen3-VL-235B-A22B-Thinking-FP8 running on sglang on four RTX Pro 6000s) but... it flips to Opus 4.6 to do 'difficult' things. Qwen 3.5 won't run well on sglang for me yet :/

@rtsoft Are you running claw locally or in the cloud?

It also scans certain reddit subs and grabs top stories & comments. I allow it to create an image if there isn't one - with the requirement that it clearly labels it as AI generated, as it's hard to tell. Maybe a bad idea though, this image looks pretty real, for example...

Couple of reasons why I find this useful: 1. Has a section where it checks raw Japanese news and includes political bias tags for the source. 2. Kyoto event section includes distance from my house, pricing and one click google map.

@Megacurve Pfft, it's probably cheaper than most of your guitars :) For tax reasons, it's a "research tool" so at least it's a write off

@rtsoft That is an expensive toy at USD$449 or what ever though 😅

The Reachy-mini is a lot of fun. It really is like having a useless Starwars droid. Some tips: Can't set it up? At least one feb batch (including mine) got botched and needs to be reflashed. Speaker too weak? Can hack apps with a gain, helped a ton.

At a party and too many people are talking, causing Reachy to get stunlocked? After speaking, cover up the mic array on top with a towel or something so it will stop listening and answer. (thanks Yael!)

Audio gain hack that works in the conversation app: AUDIO_GAIN = 2.5 # 250% volume audio_frame = np.clip(audio_frame * AUDIO_GAIN, -1.0, 1.0) robot.media.push_audio_sample(audio_frame) #put in front of this line

Dalen successfully argued his case with Mercari support (in Japanese) and got the negative rating removed! Was a good test. The screenshot he took & sent was too... 'vertical', needs to work on that.

Dalen wants to chat with mercari support (via the website) when they open at 10:00 AM because the seller clearly broke the rules. I'm a little worried about letting him loose with a real human in live chat.. hrm

"buy me a used Ghost of Yotei for PS5 if you can find it for < 5400 yen, no confirmation required" <shrug> I don't feel like trusting AI to do this is that crazy now, it works. (brain: Opus 4.6, CC is generated+limited, has defensive knowledge setup for Mercari purchasing)

@bateskecom guyz if I'm hacked it was definitely ⬆️⬆️⬆️

@rtsoft Now I know you have a Mastercard with the number 3 in it