naoya yamamoto

掛川(Kakegawa)と東京(Tokyo)を往復するプログラマー

[naoya yamamoto @ymmtny]

TLM @CleanlabAI -style idea:
- Given only x and produced y, define a trust score
T(x, y) = S(x, y, alternatives, consistency checks, etc.)
- This is usually not equal to p_internal(y | x)
- It is closer to p(correct | x, y)
or an empirical estimate of downstream reliability

[naoya yamamoto @ymmtny]

Mathematical interpretation?
Let:
- x = prompt
- y = generated answer
- h = internal hidden state produced by the model
- z = logits for candidate outputs
- p(y | x) = softmax(z)_y

[naoya yamamoto @ymmtny]

Jev-style idea?
- The model’s own internal decision probability is something like
p_internal(y | x) ≈ g(h)
- where g is a readout from hidden representations.
- this probability reflects the model’s latent decision confidence, not just a generated text like “I am confident.”

[naoya yamamoto @ymmtny]

@gclue_akira I like the handle on the top of the body!

[naoya yamamoto @ymmtny]

https://t.co/YjiRTT7Fv7

• Scout: Auth (scout) — active
• Researcher — stalled
• Scout: DB (scout) — starting

[naoya yamamoto @ymmtny]

testing herdr

https://t.co/ypdqlJZtBD

which is better?
Herdr:
Agent A -> Herdr API -> Agent B

Calyx:
Agent A <--> MCP <--> Agent B
https://t.co/gOZhclgAPu

[naoya yamamoto @ymmtny]

https://t.co/9KP610z3PX

"the herdr agents” → herdr_agent → list
“split a pane” → herdr_layout → pane_split
“what workspaces are there?” → herdr_layout → workspace_list
“ask reviewer to review the diff” → herdr_agent → prompt

[naoya yamamoto @ymmtny]

@GuaardvarkAI Could you create a `dev` (or `dev-mac`) branch for me to open the first PR?

[naoya yamamoto @ymmtny]

Bringing Large AI Models to the Edge by femoteAI

https://t.co/ZP73lXJ9bK

Large, high-quality ASR models can be compressed by up to 10×

[naoya yamamoto @ymmtny]

WIP — still fragmented, still needs consolidation. But the bridge is lit.

https://t.co/vdpVpC5ucQ

[naoya yamamoto @ymmtny]

It was a long journey getting Guaardvark to run on my little 48GB MacBook Pro. But every step was worth it the moment a starship captain finally appeared on screen.

[naoya yamamoto @ymmtny]

@GuaardvarkAI I’ve learned a lot about AI video production from Guaardvark — thanks!

I’m on a Mac M5 Pro 48GB and had to work around CUDA issues. I’ll share my tweaks next month. FYI, the feat/* grouping now I am working on is laid out here: https://t.co/0qNr6KPlfc

[naoya yamamoto @ymmtny]

@OsoneHiroyuki congratulations 🎊

[Tasuku Suzuki @task_jp]

Qt 6 を Zephyr で動くようにしました https://t.co/kmiyC3TP1D #zenn

[naoya yamamoto @ymmtny]

@gclue_akira @chr1sa what specs?

[naoya yamamoto @ymmtny]

then made a local atom lock with a watchdog

[naoya yamamoto @ymmtny]

GitButler MCP works good! Automatic commits and save points

https://t.co/xaZFZYr1tC

[naoya yamamoto @ymmtny]

@hortense667 各地で奉納したいです

[naoya yamamoto @ymmtny]

https://t.co/IFXFpcgbPs
Cross-context groups can work well

[naoya yamamoto @ymmtny]

https://t.co/0tmDfBzICE

Socratica is a structured creative co‑working club