Matt (@mko4444) header
Matt (@mko4444) profile picture

working on something new

San Francisco, CA
Joined November 2024
I had way too much time on my hands during covid... (no AI used here)
4
215
Why didn't Apple just put FaceID on the iPhone Duo to begin with? And why is it not on all of their devices for that matter?
79
WE DID IT! Esmeralda just secured its final approval, with a unanimous 4-0 vote! Approvals of this magnitude are extraordinarily rare in California. This has been a long, intense journey I am proud of our team's tenacity. Most of all, I'm deeply grateful
209
129
69
2,179
187,962
there's a reason that most live AI voice demos are of a customer service call: 1. they're short (quality breaks down over time) 2. they stay on script (usually dumber than text models) 3. they have a clear ROI (live models are expensive) IMO, the reason there isn't as much experimentation in consumer products that use live voice is because these trade-offs are hard to overcome, particularly for a small team who can't subsidize the cost like a lab.
2
6
299
teapawt.org/joshm/status/212208145… tried this demo, it's a really neat idea. very much agree with the theses that a) chat is not the universal (visual) interface and b) that talking it out via voice is better than writing. but if I'm editing a doc with GPT live (not sure what they're using but assume it's that) for 30 mins, how much does that call cost browser co? $1? $2? how do you make the economics work out there for a consumer product?
The @browsercompany has been prototyping at the intersection of (a) full-duplex voice models & (b) traditional software GUIs. “Chat” just feels too limiting for most tasks. We stumbled upon something interesting & wanted you to be able to play with it. Demo below (audio on):
1
49
Why do car companies keep adding screens that stretch across the width of the car? They look terrible.
Mercedes has unveiled its new "point-to-point" driving assistance system that it will launch in the U.S. in 2027, allowing vehicles to navigate from start to destination, including through city streets. The system can: • Turn at intersections • Stop & accelerate at traffic lights • Change lanes and merge into traffic • Navigate highway exits • Follow a route through dense urban traffic The system uses 10 cameras, 5 radar sensors, 12 ultrasonic sensors. The required hardware is already installed in certain eligible 2027 vehicles. Mercedes is also launching hands-free highway driving at speeds up to 85 mph. Select California customers will get early access to both starting next month, with a broader U.S. rollout in 2027. Video below of this new MB.DRIVE ASSIST PRO system being used in San Francisco:
2
107
Introducing AgentPaint, an oil painting environment in the browser for AI agents (and humans) We've seen how great frontier models like Fable 5.2 are at painting directly in code. AgentPaint takes a different approach to exploring models' visual and creativity abilities, by giving them control of a paintbrush in the browser, on a digital canvas designed to simulate physical oil painting. Agents can only paint by moving the brush on the canvas, and they can only get the colors they want by mixing existing paints. They can't manipulate pixels directly or render the painting as code. Here's a few examples of different models painting the same prompt in @AsideAI (no affiliation, it's just a great model-agnostic browser/agent) /1 Van Gogh's The Starry Night, from memory
5
4
30
1,447
This moment in history should be required viewing in every school in America.
2,987
25,455
1,386
114,522
4,220,150
saw god today – at California, USA
35
63
8
891
77,025
fucking insane wow
Mechazilla catches Starship! Congratulations to the entire SpaceX team on this incredible achievement!
1
47
It’s like dark mode: you only focus on it once the core stuff is covered. OpenAI is clearly not on their game atm. So yes, them focusing on this right now would get (correctly) roasted.
it's crazy how the reaction to a feature launch from an AI lab is COMPLETELY dependent on how good their current models are if OpenAI launched this feature they would've been crucified in the comments "where's the models" "astra 6.2?" "still mogged by opus 5.5" "models??"
2
148
If it’s comparable to opus on intelligence and personality but half the price (or ideally even less), I don’t see why devs wouldn’t build on it over opus
GPT-6.2 Sol will be the most important release for OpenAI, and the odds are stacked against them. They are up against a monster of a model aka Opus 5.5, that hits the trifecta of speed, cost and intelligence, and matching that would already be a feat on its own. And even if they somehow managed to make GPT-6.2 Sol better, chances are it’s marginal, and would most likely be mogged by Opus 5.6 soon after which had a 2 month head start. The only way they can come out on top is if they beat Claude models by 2 generations in 1.
2
144
Still blows my mind how much faster I can move because of AI. Not quite at the speed of idea generation, but getting close.
1
74
We’ve open-sourced our 4B full-duplex speech model under Apache 2.0 !! Our model was trained on 40M hours and uses continuous tokens. Full-duplex AI for "Human like AI" is still largely unexplored. Let’s push the boundaries together! Check out our demos! 🚀 Website: higurashi.ai/haru4b/ Code: github.com/higurashi-ai/Haru Model: huggingface.co/higurashi-ai/… Technical blog: blog.higurashi.ai/melte/haru…
【人間のように、自然に笑い、相槌を打つ「人間らしいAI」】 4BのFull-duplex音声対話モデルを4,000万時間の日本語、英語を含む多言語データで事前学習し、オープンウェイトで公開しました! 弊社のオープンモデルを用いて、開発者の方々と一緒に未来を切り開ければと思います
13
125
11
999
57,581
AI interfaces nowadays are still pretty much turn-based. In my PhD thesis, I thought about what interaction would look like if we could go beyond taking turns. One example is intra-turn interaction:
306
616
220
8,577
805,004
Update: I started a new company! As AI makes it easy to remix any content, a new type of role is emerging: the media engineer. Their job is to build systems that continuously turn a company's work into a steady stream of media.
100
16
11
1,406
36,874
interesting that they took the opposite approach of all the other products in this vertical. Granola/wispr/notion/chatgpt all join silently
Claude can now join your meetings as an active participant. It’s unlike anything I’ve used on a call, it can: • Build a doc or whiteboard live as you talk • Check connected tools with your OK • Answer the team when asked out loud The work kicks off before the call ends.
1
1
216
All the AI voice demos are cool, but live models are still super expensive to run in prod, and the quality drops the longer the convo lasts. The more exciting area is the cascade stack. You can now run a ~realtime STT-TTS pipeline on your iPhone for free, without sacrificing performance or battery life. The drop in cost opens up new product ideas that would not have been viable before. And, it makes it far easier to add voice to an existing agent without needing to rewrite your harness around a live model's black box.
3
146
San Francisco just gave final approval to the controversial 848 home Marina Safeway project. Here's the approved plans: • Two residential towers, 18 and 22 stories • 258 feet tall • 848 apartments • 86 affordable homes • 67,180 SF of space for a new Safeway • 349 car parking spaces • 360 bicycle parking spaces • Resident amenities and open space Right now it's a Safeway and a surface parking lot on a 2.6-acre block. It got fast-tracked under AB 2011, and it actually grew along the way. 22 stories in the Marina. Didn't think I'd see that one 😅 📍San Francisco
67
26
10
575
65,257