Principal Designer @ https://xitter.cf/t.co/7kkTWHSRpe (NYSE: XYZ) Currently building Buzz https://xitter.cf/t.co/74AvLgvc46 Views are my own.
New York
Joined September 2006
- Tweets18.5K
- Following115
- Followers5.9K
- Likes11.2K
ThomPete retweeted
i think this work is interesting but i don’t believe that this is what people want.
sci-fi per usual has forecasted the end state.
HAL, C3P0, Data, Jarvis, and whatever that thing McCoughnahey was talking to in Interstellar. way better targets than mimicking a human.
ThomPete retweeted
I'm happy to announce that Veritas Kanban 6.3.0 adds support for signing in with ChatGPT by linking your OpenAI identity to your existing VK account.
If you’re using AI agents for work, VK gives you a place to keep track of what you’ve asked them to do. Put the work on a Kanban board, assign an agent, and keep the instructions, approvals, and results with the task. When you come back later, you can see what ran, where it stopped, and what needs your attention.
You can use different agents for different tasks and set up recurring workflows. OrcaRouter support is new in this release too, giving you another option for routing work through Codex CLI.
A lot of the work on 6.3.0 was making that day-to-day experience less frustrating. Moving between tasks without losing edits. Getting the board updating again after a dropped connection. Knowing whether an agent is working or waiting. Having scheduled work run when you approved it to run. There are security, storage, and Mac app fixes behind those improvements, with the details in the release notes.
VK is local-first, and the signed and notarized Apple silicon app is available through Homebrew or direct download. The updated docs cover setup, including the new integrations. Check out the full release notes and download:
github.com/BradGroux/veritas…
ThomPete retweeted
A 4B parameter model with strong agentic and long context capabilities, running in your pocket, on an @Apple iPhone.
ThomPete retweeted
We built our own mega kernel @axioniclabs as well last week!
Runs all on a single RTX 5090 while enabling our custom model architecture to learn, generate, and actually play (and improve on) Doom.
Will love to collab!
ThomPete retweeted
Block（ジャック・ドーシーの会社）の会話モデル「Dessa」えげつないな。
文字データも事前学習済みモデルも一切使わず、人間の音声だけで訓練したらしい。聞きながら話せるのでサンプル音声も、もはやどちらがAIか区別できない。しかもこれ、MacBook Pro M4 Max以上ならローカルで動くらしい…
Last week, @blocks introduced Dessa, probably the most realistic voice model you've ever heard.
80% of the world's workforce doesn't sit behind a computer to do their jobs, many of whom we serve. They are the retail workers, the restaurant workers, the plumbers, the hairdressers, the service technicians, the truck drivers, the park rangers, and the builders.
For these people, conversation and verbal communication are just as important an interface, if not more so, than they are for information workers.
We built Dessa because we believe in creating meaningful relationships with technology by adapting to how humans work, rather than forcing everyone to adapt to technology.
Dessa is full duplex, meaning the model listens while it talks, and it's trained entirely on voice (i.e. no LLM).
On top of that, it can actually run locally on your own computer and, eventually, your phone.
It's currently available to researchers for testing and analysis.
Check it out...
ThomPete retweeted
来源：@jack 的发布帖 xitter.cf/jack/status/2113697902…
Block 团队的补充说明 xitter.cf/Hello_World/status/211…
提醒家里人：接到"熟人"语音借钱，先挂断再回拨确认。
Last week, @blocks introduced Dessa, probably the most realistic voice model you've ever heard.
80% of the world's workforce doesn't sit behind a computer to do their jobs, many of whom we serve. They are the retail workers, the restaurant workers, the plumbers, the hairdressers, the service technicians, the truck drivers, the park rangers, and the builders.
For these people, conversation and verbal communication are just as important an interface, if not more so, than they are for information workers.
We built Dessa because we believe in creating meaningful relationships with technology by adapting to how humans work, rather than forcing everyone to adapt to technology.
Dessa is full duplex, meaning the model listens while it talks, and it's trained entirely on voice (i.e. no LLM).
On top of that, it can actually run locally on your own computer and, eventually, your phone.
It's currently available to researchers for testing and analysis.
Check it out...
ThomPete retweeted
We actually explored what makes Dessa feel so human, which matters for your 80% workforce point here: xitter.cf/AGTPinsights/status/21…
Block just unveiled Dessa, a speech AI model trained from scratch on nothing but speech. Here's what you need to know.
Ragavan Thurairatnam of Block announced it on X today, Oct 24, 2026. His pitch: speech is the natural interface for technology, so Block built Dessa. Block's own post calls it a speech model trained from scratch on human conversation. The company's page calls it the Conversation Model trained only on speech.
Dessa is full duplex. It listens while it speaks. Early listeners say it understands timing, emotion, laughter and pauses. One called it the most real AI voice they've experienced. Another called it the GPT-3 moment for speech.
Block is also inviting AI safety researchers to explore Dessa. The blog post has more samples to hear. Block's Jack Dorsey posted "new from us" alongside the launch. One team member said an app is coming soon.
Key numbers:
- Trained from scratch on speech only, no text pretraining
- Full duplex, real time speech to speech
- One early commenter says it was trained on a single GB200 NVL72 rack, not confirmed by Block in the material
- Launch demo video runs about 3 minutes 27 seconds
Some reactions: people want more realtime voice competition outside OpenAI, and one researcher hopes the weights get opened. One viewer found the voice eerie. Block hasn't announced pricing or a public release date yet.
ThomPete retweeted
running locally is the part that matters for a plumber or a clinic, every customer call going to someone elses server is a data decision most of them never made
