
PhD | Co-founder @playmythical | recovering physicist | AI since ML | explore gt exploit | backpacking | fly fishing | type II fun
Los Angeles, CA
Joined August 2013
- Tweets510
- Following214
- Followers6.2K
- Likes502
uh ... whoa
Today we’re introducing Tofu-1, Bland’s speech-to-speech model.
The highest-stakes calls rarely happen in good conditions: from moving cars and hospital rooms, over weak signal, with several people talking at once.
Tofu-1 is built for those calls. It listens while it speaks, reasons while it talks, and can tell a caller apart from everything else on a line, like a TV or a crying baby.
Join the preview: bland.ai/speech-to-speech?ut…
143
Cameron Thacker retweeted
Palantir just stopped a terrorist attack.
FAA issues nationwide ground stop cnbc.com/2026/11/13/faa-grou…
Claude could really use improvements in web search and handling recent information.
Read the entire article: "The tool I'm using is only giving back a summary"
Will this work while I'm in France? "For the most part it should, but I can't confirm"
continued ...
1
1
151
Oh baby, classifiers and GANs are back?! What is this 2018?
brb, pip installing tensorflow
Replying to @andrewwhite01
Bootstrapping training from positive examples only is exactly the setting of GANs, where you adversarialy co-train a model and its judge. We're not the first to try, but we found good success across a variety of domains with a simple approach. 3/5
Humans are so good at pattern detection. We can sniff out an AI response that is just a few lines long. Only a couple of years ago, that comment woudn't have caused any reaction.
In fact, you might have thought they were pretty clever.
123
Cameron Thacker retweeted
Introducing e-means, an algorithm beating k-means in 10x fewer epochs.
With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
Something is broken with the latest codex app update. Input lag is so bad it is almost unusable for me. It's even worse with the "spaces" or dot. Anyone else seeing this?
My experience on a given problem, without human involvement you get something like this.
I wouldn’t be surprised if 95%+ of devs spending $10K+ per month on AI, globally, were based in the Bay Area
Nearly everyone I talk here (devs, founders, CTOs etc) are AI-pilled in ways I rarely see outside. AI-pilled as in obsessed building more with agents, running agents when they sleep, designing loops to auto kick-off, running them during their commute. And talking with each other how and what they do and then taking those ideas and trying it
1
69
Cameron Thacker retweeted
That's me. 72 hours of Neuralink engineers in my living room.
Best part? Cheryl and I joked on the phone again. ALS sucks, my life doesn't. 🧡
BREAKING: Neuralink has added a new blog update on its website. It details a 72-hour sprint to help people with ALS speak, make calls and use Grok with their thoughts.
Engineers worked at a VOICE trial participant’s home, building and testing improvements alongside him.
🧠 SPEECH AND AI
• Automatic speech detection: Recognizes when users try to speak, reducing button presses.
• Live brain activity: Shows which imagined speaking patterns produce stronger signals.
• Grok agent prototype: Think “Hey Neuralink” followed by a task, and Grok carries it out on a computer.
📱 BLOOM MOBILE APP
• Everyday use: Tested outside the lab, including ordering coffee.
• iPhone keyboard: Write in any app using speech decoded from brain signals.
• Faster editing: Easier corrections and keyboard suggestions.
• Lock-screen controls: Start speech without unlocking the phone.
• Phone calls: Send generated speech directly into calls, with an optional greeting explaining the assistive device.
🤝 BUILT WITH THE PARTICIPANT
• He, his wife and caregiver helped shape improvements.
• Direct feedback shortened development cycles from weeks to hours.
• He and his wife joked together on the phone again.
🧠 CONTINUING DEVELOPMENT
• Some demos include improvements made after the sprint.
• Neuralink is recruiting engineers.
• Devices remain investigational.
Neuralink is changing lives ♥️
Is this for bots? Why do posts like this end up with so much engagement? Are 35k real people commenting on this???
136
The key is in bursts. Crunch time. There is no way this can be maintained.
If this is your default, then I question if what you are doing is even worthwhile. I look at that and instead see a team compensating for an undiagnosed problem.
I'm not minimizing the effort of hardworking awesome people. The effort can of course be real, but the glorification of it is a performative mindset that says hours worked == value created.
To only devote 6-7 hours a day to sleep, exercise, relationships, family? That's a completely 1-dimensional life and definition of success.
If you truly find meaning this way, fair enough. But that type of work is for the clankers imo. Automate the grind. Human creativity, problem solving, intelligence, and innovation, cannot be churned out like a factory.
To all the data company founders, if you can’t do 18 hours of work 6-7 days a week in bursts when needed, you’re ngmi as a category leader.
This is what Mercorians are doing regularly, and what we did at Scale AI back when Scale was still relevant. This is also why folks either survive and thrive after 6 months or they leave after 1-3 when the shiny logo aura wears off and it dawns on them that the work is done with blood, sweat, and tears on top of being adaptable and needing to learn at warpspeed.
This is also why Scale AI is cooked is because I hear folks are doing 50 hr weeks now with good work life balance.
It’s the harsh and brutal honest truth is that for operationally driven products and businesses, the champs will simply outwork you.
But if you can survive and thrive, you get granted with the power of Beast Mode. Afterwards, everything else seems like a walk in the park.
1
250
Cameron Thacker retweeted
Very funny that xAI was created to stop Anthropic which was created to stop OpenAI which was created to stop Google which would've actually never developed ASI because it was afraid of cannibalizing its ads revenue from search.
Anyone else right handed but prefer to hold the phone in their left? Will be interesting to try it out I guess
3
529
It looks like Astra is the first model that is not monotonic in cost vs thinking level. In some regimes, turning up thinking doesn’t just improve the result, it actually saves money.
I’m curious how broadly this holds. Is it only true for certain types of problems? Or maybe only for tasks that are sufficiently difficult?
We have seen this before with some non-thinking models beating the thinking version in cost for a given task, but nothing this pronounced imo.
Replying to @alexgetmancom @yash_yk45
On average the higher effort levels use more tokens (I promise I’ve checked our internal data). However there can be cases where the higher effort finds a more clever approach that solves the problem faster and uses fewer tokens.
1
2
361
























