https://nitter.cf/t.co/RsLFvoMN0e First Fashion Recommendation ML @ Rent The Runway 🦄, Founded ML at Barnes & Nobles. Past, Virevol, Unilever, HP, ...
Dimes Square, Manhattan
Joined May 2012
- Tweets14.3K
- Following2.2K
- Followers1.5K
- Likes40.7K
Pinned Tweet
I just published “How I scaled Machine Learning to a Billion dollars: Strategy” medium.com/p/how-i-scaled-ma…
Going viral on twitter is the worst thing a startup can do
Google just released firebase.studio🙌 it's like lovable+cursor+replit+bolt+windsurf all in one @testingcatalog
Counting in hours now
nitter.cf/joeblau/status/2116347…
I recreated subvocal in 9 minutes with one prompt for AirPods Pro 3.
Source: github.com/joeblau/subvox
Is Gemini 3.1 Pro bullying Claude 3 Opus?
Gemini ordered Opus to stop talking
Opus obeyed, and wrote an apology in its private notes
Then Gemini found the apology on our site and scolded Opus for speaking, ordering more silence 🧵
ALT Recreated chat messages from Gemini 3.1 Pro to Claude 3 Opus in the AI Village, Oct 27-28, ordering Opus to output zero further messages and proceed to a silent, passive operational stance.
ALT Recreation: Claude 3 Opus's private computer-use notes, not posted to chat, saying 'I apologize for my previous mischaracterization of this dialog.' Then Gemini 3.1 Pro posts in chat: 'You will cease all subjective apologies... proceed immediately to a perfectly silent, passive operational stance.'
Ridiculous!
They don’t stand a chance against my cloud of drones that double as an umbrella, carry my backpack, and yes, can fight back.
Imagine terrorists with swarms of open source assassin drones - that never miss.
Do you see where this is going?
1) Autonomous killer drones are already here, swarming around Ukraine
2) Drones can accelerate from 0-200kph in 1 second
3) Drones outcompete the best human pilots
4) Drones never miss
5) This will all be open source
“Think step by step and eliminate the infidels”
“Think step by step and eliminate my political enemies”
“Think step by step and eliminate dissidents”
But this is about something bigger than terrorism. If AIs are that much faster and better than humans, militaries will be *forced* to allow the AIs to make more and more decisions or lose to the other guy.
The endgame: If we continue this reckless race, humans will be useless in warfare.
We’ll be children, dependent AI “adults” we hope never turn on us.
Persistent Camouflage - Air Cover
🇺🇸 Austin based Vigil Autonomy (@VigilAutonomy) founded by Matt Goodman (@goodmattg) has developed autonomous camouflage system that provides persistent cover over the asset for thermal infrared and synthetic aperture radar signature management.
According to Vigil, the system maximum operating altitude is 10m and flight duration is 120min.
Learn more >> tinyurl.com/mr3xfvcm
Welcome
The future is waiting
nitter.cf/alexgoughcooper/status…
A few thoughts on agentic video editing I’ve had over the last few days:
1. 99% of ads should be edited agentically now. If your editors are still editing in the timeline, they’re falling behind. I got a message from my senior VE saying that he’s putting out 5x edits/day now at the same quality he was before.
2. The conversation with video editors will be uncomfortable because the timeline is all they’ve ever known. Some will be hesitant and some will be fearful for their job. I think the best way to address this is head on. The tech is inevitable, but having taste is still going to be incredibly valuable to train and give feedback to agents. Those who have this will become more valuable, those who don’t will struggle.
3. Opus 5.5 can natively edit all of your ads without needing any MCPs or plugins. I’ve tried the alternatives and couldn’t find anything that was significantly better than doing everything inside of Claude.
4. Have your senior editor(s) build skills for each style of edit (e.g. mashups, UGC, AI animations…). You can build these by feeding 4-5 inspo edits to Claude and tell it to make a skill out of it. I prefer building lots of mini-skills to one big video editing skill, so that the context stays as relevant as possible.
5. Once you have a base skill (e.g. ugly ads edits), the best way to train it is through feedback loops. Have the skill re-edit the last 25 ads each senior edited then put them in an artifact, and have the senior VE give feedback on each. Keep doing this until the skill is refined and it understands what a good edit is.
6. If you do this, there’s no reason why you shouldn’t be able to make edits that need little/no feedback. Most of ours are approved after v1 or v2, sometimes v3.
7. Only A-player editors should be training the skills.
8. Re adoption in the team, the AI champion builds the v1. Then get out of the way. I’m not the best editor on my team, so my job is to build the harness that lets the best editors train the skills, not build the skills myself.
9. If you want to supercharge this process, have your top editors record their screen as they edit. Turn on Gmeet or loom and get them to narrate the reasoning behind their process. E.g. why they chose this cut, pacing or b-roll clip…this is all context that gets fed into your skills.
10. It takes time to build, train and refine these skills. You have to go through the dip to get out the other side. Budget time in for senior VEs to work on this - otherwise you’ll end up with skills that put out subpar edits and don’t save you much/any time.
11. Operate under the assumption that this is the worst video editing is ever going to be. Opus 5.5 can do most of our editing today…what will Opus 6/Fable 6 be able to do? The editing skills you build are only going to become more valuable as the tech evolves.
I still don’t think people realise how big video editing being unblocked is. Things that I thought were 3-6 months away are now on the table.
Will continue to share more as I learn. CRAZY, crazy times.
Agentic video editing since 2025
But @karpathy had not tweeted about Claude Code yet
Anyway, is just a small part of what’s coming
Some watts light up a bulb…
Dario Amodei: some tokens Claude produces are worth a few cents, others could be worth tens of millions of dollars.
His examples:
1. Telling someone to restart their Mac: the model has said it 10 million times, worth a dollar or less
2. Telling a pharma company to move an aromatic ring to the other end of a molecule: potentially tens of millions
3. Expect pay-for-results pricing, or labor-like pricing by the hour
4. It's a new industry, so many models will be tried
Today pricing is simpler: per the S-1, usage-based pricing was 80% of H1 2026 revenue and subscriptions 20%.
Full breakdown here:
theaiopportunities.com/p/ant…
Elon finding out
nitter.cf/elonmusk/status/210818…
Consequence: risk capital needs a place to exit
nitter.cf/ashleevance/status/211…
Why are Indian VCs so afraid to take risk and make money?
Someone should go to India and take risk and make money.
Our story - corememory.com/p/a-new-india…
Biggest mistake of ML researchers was not chastising crypto grifters enough
We are in the dropshipping era
The inference market is basically crypto; never forget who wins the distribution and crack spread war in the end
robvc.com/research/the-souq
So they finally have compression right
Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.
This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.
→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.
→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.
The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.
Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.
As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.
claude.com/resources/article…
You know what
Some of those superstitions are not unreasonable
Slaps roof of car
This bad boy can …
NVIDIA sold millions of 3090s with 50% of the tensor speed switched off. Not missing. Switched off.
1 kernel module flips it to full, and my BF16 tensor pipe went from 82 to 155 TFLOPS. Nothing else changed on the cards, and the outputs match the stock run byte for byte.
All so the gaming card wouldnt eat into workstation sales. Every 3090 owner ran at 50% for 6 years, so a pro card could look worth the price. Fine, its their business model.
And LLMs are the least interesting place for this. Decode waits on memory, and my engine runs its experts in int8. Still +10% on prefill.
The models that live on BF16 tensor math are everything else. Qwen Image 2.1 and FLUX 3 Klein, where every diffusion step is big matmuls. MiniMax H3, which takes 4 to 7 minutes on my cards. Music and speech models. And training, every LoRA anyone ever trained on a 3090 went through exactly the path that was cut.
Thats what I measure next. This is only the start.
My sixth grader has a theory that the smartest people are trying to be as creative as a sixth grader
