Pinned Tweet
UK ROGUE AGENT RESPONSE | QWEN NEURONS | TOUCHSCREEN MACBOOK xitter.cf/i/broadcasts/1dxYlaZzE…
MTS retweeted
LIVE AT 1:30 PM ET — FCC National Security Counsel joins @MTSlive
Don’t miss it👇
UK ROGUE AGENT RESPONSE | QWEN NEURONS | TOUCHSCREEN MACBOOK xitter.cf/i/broadcasts/1dxYlaZzE…
MTS retweeted
we covered the largest robotics seed round ever raised in the uk, on mts industry news today
congrats @Machenta_ on the big day, and thanks for coming on with such short notice (booked 5am pt!)
Machenta CEO @Ed__Johns explains how his team automated an entire assembly line in just 2 minutes, something that normally takes weeks or months:
"I decided I wanted to automate an entire assembly line. We put three robots side by side and created a task where the robots would package a box and then close the box, and each robot would do a different part of that task."
"I thought it’d be really cool if we could automate that whole thing in about two minutes, because traditionally automating an assembly line would take weeks or maybe even months of systems integration and manual programming. It’d be extremely expensive."
"If we could get instant robot learning working on an entire assembly line, this would be evidence of what the future of the manufacturing industry would look like."
"That’s one of my favorite tasks, an entire assembly line automated in two minutes."
@Machenta_
MTS retweeted
awesome to have @general_compute on industry news today. congrats on the $100m round!
General Compute CTO @GoodisonJason predicts AI is about to have its broadband moment, making today's slow AI agents feel as outdated as dial-up internet:
"We’re in the dial-up phase of AI right now, where people think it’s normal to run a few agents at a time and go for a cup of coffee while something is running. You wouldn’t start loading a website and walk away."
"We’re moving into what we like to call the broadband moment of AI, where applications actually have to be interactive and fast."
"It hasn’t really been brought to open source models, and it hasn’t been put in the hands of builders and founders to build completely new product experiences."
"We’re the neocloud that actually buys the hardware and deploys it physically."
@general_compute
General Compute CTO @GoodisonJason explains why GPT-4 conditioned us to believe slower AI means smarter AI, and why he'd now choose speed over intelligence:
"Generation speed used to be a proxy for how smart the model was. GPT-3.5 Turbo was really fast, and then GPT-4 was really slow. I was okay to have it be slow because I at least had this proxy of, that’s a smart model."
"Now open source models are catching up to the frontier. Many tasks don’t need the frontier anymore. So now you’re gonna start seeing people demand speed."
"Even if the response was slightly worse, I would rather the fast information, because I know it already has sufficient intelligence to get the task done."
@general_compute
🚨 Big news today: We’ve raised $100M for our Series A! 🎉 🚀
We believe demand for fast inference is many times larger than demand for slow inference, and it’s about to go exponential.
Thanks to our co-leads @BlackbirdVC and @Abstractvc. Two visionary partners join our journey 🎉 🚀
General Compute CTO @GoodisonJason predicts AI is about to have its broadband moment, making today's slow AI agents feel as outdated as dial-up internet:
"We’re in the dial-up phase of AI right now, where people think it’s normal to run a few agents at a time and go for a cup of coffee while something is running. You wouldn’t start loading a website and walk away."
"We’re moving into what we like to call the broadband moment of AI, where applications actually have to be interactive and fast."
"It hasn’t really been brought to open source models, and it hasn’t been put in the hands of builders and founders to build completely new product experiences."
"We’re the neocloud that actually buys the hardware and deploys it physically."
@general_compute
🚨 Big news today: We’ve raised $100M for our Series A! 🎉 🚀
We believe demand for fast inference is many times larger than demand for slow inference, and it’s about to go exponential.
Thanks to our co-leads @BlackbirdVC and @Abstractvc. Two visionary partners join our journey 🎉 🚀
Dirac CEO @FilArons warns that AI robotics startups raising millions could repeat the same mistake that killed robotics companies a decade ago:
"When I see robotics companies raising a ton of money to deploy robotics into a manufacturing environment, I think we’re seeing a misunderstanding of what problems AI actually solved. We did not solve the integrator problem for robotics."
"There were a bunch of robotics companies in 2016, 2017, 2018 that ended up dying because, as it turns out, as a robotics company, you are functionally a robotics integrator. It’s really gritty, hard work that most tech folks don’t wanna do."
"We work with some legacy defense products. They build missiles to keep our country safe. A lot of these folks are just now getting on the cloud. The right way to do it is just meeting them where they are."
@DiracInc
Dirac CEO @FilArons explains why a welder with 20 years of experience could beat a Stanford graduate at bringing AI to manufacturing:
"A lot of what we’re seeing is people throwing AI at legacy manufacturing problems."
"It’s directionally correct, but I’m not seeing a ton of focus on solving specific problems. Not just building solutions looking for problems."
"These folks who actually know about the core problems in Western production are getting really good at AI."
"Who’s going to implement AI in these legacy production processes? Is it gonna be the Stanford kid at 22 jumping into a mom-and-pop shop with AI and some finger guns? Or is it gonna be the dude who’s been welding for 20 years just solving his own problem?"
@DiracInc
Standard Bots CEO @evanbeard reveals why nearly 99% of factory robots still operate like arcade claw machines, and how AI could finally change that:
"Almost 99% of manufacturing robots today, people don’t realize, they just are replaying the same motions again and again."
"You’ll save points, kind of like the crane game at Denny’s, and then the robot will replay through those points. That’s all those robots really do all day long."
"We’ve released a vision encoder trained on billions of images where you can really quickly train a model to find objects and pick them up, and few shot learning, where you demonstrate a task to the robot and it learns from you."
"One is making the robots so anyone can take them out of the box, like an iPhone type experience for what traditionally is a mainframe computer type product."
"The other is making them more capable, so you can automate jobs that previously had too much variability."
@standardbots
Introducing Standard Bots AI
“We chose Standard Bots for two reasons: the speed of development is highly compressed, and the interface for programming the robot is so much simpler. With Standard Bots AI, that's even faster.” — Adrian Nistler, President and CEO, Path Machining & Automation
Standard Bots AI robots deliver dramatically faster setups, faster changeovers, and faster paybacks. This is why robots are the new essential power tool of the 21st century.
Our mission is to expand human capacity with the world’s most accessible robots. We’re building this future in Glen Cove, New York, and on hundreds of factory floors, and counting.
Say yes to more: standardbots.com/ai
Machenta CEO @Ed__Johns reveals how betting against the robotics industry's approach led his team to build robots that learn new tasks instantly from just one demonstration:
"I decided what I wanted to build is a robot brain that can learn tasks like a human. And the rest of the world went off in a different direction."
"They started collecting lots and lots of data in the real world, with the assumption that you would need to provide hundreds of demonstrations."
"I always thought there’s something not quite right about that, where a robot would need to spend hours or days to learn a task. That’s what I bet on at Imperial College."
"The internals of it are nothing like the human brain. But the way you interact with it is like interacting with a human."
"If I wanted to teach you a task, I could give you a demonstration, and you’d be able to do that task. You wouldn’t need hundreds of demonstrations, and you wouldn’t need days or weeks of training."
"You provide one demonstration of the task, and then instantly the robot will be able to perform that task. Which is very different to what’s been going on elsewhere in robotics, where the assumption has been that you need lots and lots of data to teach a robot even a basic task."
@Machenta_
Today we’re out of stealth, and we’ve raised a $40m seed round co-led by @creandum and @usv, to bring Instant Robot Learning to life.
For the past few months, we‘ve been busy building a new company.
Machenta.
This follows results from a new kind of robot foundation model we’ve built, which I’m so excited to share with you today!
Usually, it takes hours/days of data collection and training to teach a robot.
We’ve built a model that learns tasks instantly: one demonstration, no further training.
A model trained entirely in simulation, without a single pixel of real-world data.
A model where teaching robots is so quick and easy, it feels like physical vibe coding.
A model for robots that learn like humans do, from someone like you.
Welcome to the era of Instant Robot Learning.
Thank you @Kathi_Wilhelm, @rebeccakaden, @hkanji, @jonno_evans and all of our angels for your support!
@Machenta_
Learn more: machenta.com/blog/instant-ro…
See below for what you can do with Instant Robot Learning!
Machenta CEO @Ed__Johns explains how his team automated an entire assembly line in just 2 minutes, something that normally takes weeks or months:
"I decided I wanted to automate an entire assembly line. We put three robots side by side and created a task where the robots would package a box and then close the box, and each robot would do a different part of that task."
"I thought it’d be really cool if we could automate that whole thing in about two minutes, because traditionally automating an assembly line would take weeks or maybe even months of systems integration and manual programming. It’d be extremely expensive."
"If we could get instant robot learning working on an entire assembly line, this would be evidence of what the future of the manufacturing industry would look like."
"That’s one of my favorite tasks, an entire assembly line automated in two minutes."
@Machenta_
Today we’re out of stealth, and we’ve raised a $40m seed round co-led by @creandum and @usv, to bring Instant Robot Learning to life.
For the past few months, we‘ve been busy building a new company.
Machenta.
This follows results from a new kind of robot foundation model we’ve built, which I’m so excited to share with you today!
Usually, it takes hours/days of data collection and training to teach a robot.
We’ve built a model that learns tasks instantly: one demonstration, no further training.
A model trained entirely in simulation, without a single pixel of real-world data.
A model where teaching robots is so quick and easy, it feels like physical vibe coding.
A model for robots that learn like humans do, from someone like you.
Welcome to the era of Instant Robot Learning.
Thank you @Kathi_Wilhelm, @rebeccakaden, @hkanji, @jonno_evans and all of our angels for your support!
@Machenta_
Learn more: machenta.com/blog/instant-ro…
See below for what you can do with Instant Robot Learning!
Mercor's @edwardjhu reveals how AI benchmarks can reward models for guessing 10 different answers while penalizing others for getting it right:
"A rubric item might be, the final answer needs to be $3 million. Imagine the rubric item specifies it needs to be $300 million with two decimal points, and the model just said $3 million."
"Even though it’s numerically correct, because the rubric is overly specific, we’re failing the model in an unfair way."
"We released a benchmark called APEX Agents earlier this year, mirroring how a professional would actually work in front of a computer. Real professionals come up with legal or consulting tasks for the model to do over the course of 50 to 100 steps."
"One thing we noticed is that some models that are not in the top model family in terms of parameter size were doing much better than we expected. It turned out the model was doing what we call scattergunning. It will say, if A is true, it’s 2 million. If B is true, it’s 3 million. It might do it 10 times."
"As long as the model mentions 3 million, the LLM judge would say, awesome, you got it right."
@mercor
GPT-6.1 Terra is OpenAI’s #1 model on APEX-Agents.
68.3% Pass@1 (#6)
79.3% mean score (#6)
Compared to last generation GPT-5.6 Terra, the new model gains 10.1 points on Pass@1.
Management consulting: 80.9% Pass@1 (#2)
Investment banking: 66.2% Pass@1 (#7)
Corporate law: 57.7% Pass@1 (#39)
The model is strongest on management consulting tasks, ranking second overall, behind Gemini 4 Argon (88.4%).
Against GPT-5.6 Terra, it gains 24.7 points in consulting and 10.9 in banking. Law goes the other way, down 5.4 points from GPT-5.6 Terra.
GPT-6.1 Terra still averages 80.6% of rubric criteria on law tasks, but fully completes only 57.7% of them. It gets most of the way and misses the last requirements.
Across 4 runs per task, GPT-6.1 Terra passed 134 of 240 tasks every time and 187 at least once. 22 tasks it never scored on, 14 of them in investment banking.
GPT-6.1 Terra’s strength comes from reading more than its siblings: 735k input tokens per attempt, vs 497k for GPT-6 Astra and 489k for GPT-6.1 Sol. Despite reading more, it costs less.
GPT-6.1 Terra: $0.19 per task ($1/$5 per 1M tokens)
GPT-6.1 Sol: $0.26
GPT-6 Astra: $1.30
It’s also a fast model. The median attempt finishes 1.2x faster than GPT-6 Astra and on par with GPT-6.1 Sol.
Congrats @OpenAI on the launch.
See full leaderboard: mercor.com/apex/apex-agents-…
Mercor's @edwardjhu reveals why researchers are designing AI benchmarks with tasks even humans can't solve, just to make models fail:
"The benchmarks we have today are not gonna be good benchmarks in a year or two or even less."
"As the models become more capable, it’s increasingly hard to fail the model. And there is this tendency to fail the model at all costs, often at the expense of fairness or benchmark diversity."
"We’re asking the AI to do something that’s actually impossible, a human might not even be able to do it, or we find one way the model fails and we replicate it so many times"
"If the entire benchmark is testing just one failure mode, the benchmark score will swing from zero to near 100% very quickly. It’s easy for someone to look at a number and draw conclusions right away."
"But we really need to be thoughtful about the underlying design so that it’s not just a number people can cite, but a reliable instrument for the field."
@mercor
GPT-6.1 Terra is OpenAI’s #1 model on APEX-Agents.
68.3% Pass@1 (#6)
79.3% mean score (#6)
Compared to last generation GPT-5.6 Terra, the new model gains 10.1 points on Pass@1.
Management consulting: 80.9% Pass@1 (#2)
Investment banking: 66.2% Pass@1 (#7)
Corporate law: 57.7% Pass@1 (#39)
The model is strongest on management consulting tasks, ranking second overall, behind Gemini 4 Argon (88.4%).
Against GPT-5.6 Terra, it gains 24.7 points in consulting and 10.9 in banking. Law goes the other way, down 5.4 points from GPT-5.6 Terra.
GPT-6.1 Terra still averages 80.6% of rubric criteria on law tasks, but fully completes only 57.7% of them. It gets most of the way and misses the last requirements.
Across 4 runs per task, GPT-6.1 Terra passed 134 of 240 tasks every time and 187 at least once. 22 tasks it never scored on, 14 of them in investment banking.
GPT-6.1 Terra’s strength comes from reading more than its siblings: 735k input tokens per attempt, vs 497k for GPT-6 Astra and 489k for GPT-6.1 Sol. Despite reading more, it costs less.
GPT-6.1 Terra: $0.19 per task ($1/$5 per 1M tokens)
GPT-6.1 Sol: $0.26
GPT-6 Astra: $1.30
It’s also a fast model. The median attempt finishes 1.2x faster than GPT-6 Astra and on par with GPT-6.1 Sol.
Congrats @OpenAI on the launch.
See full leaderboard: mercor.com/apex/apex-agents-…
ARC Prize President @GregKamradt explains why AI researchers could be among the last jobs automated:
"AI is very good at executing what you want it to do. It’s not very good at coming up with ideas. It’s not very good at coming up with hypotheses."
"We tried to make ARC-AGI-3 games with AI-generated ideas, and they’re all crap. They lacked the special substance about what we wanted."
"On the recent podcast with Dwarkesh, he was asking, what’s the last job people are gonna have? It was along the lines of AI researcher, because AI researcher’s gonna be the one to pose the good questions."
"A really good benchmark tests how well you’re able to come up with good questions, how well you’re able to come up with good hypotheses, with an uncapped ceiling."
"If you think about what society has done over the last thousand years, there’s no ceiling to the progress we’ve had. Clearly there’s a really good uncapped benchmark out there, and it’s our job to go and find it."
@arcprize
ARC Prize President @GregKamradt explains why the final breakthrough to AGI might come from an entirely different branch of AI research:
"Our job is to find benchmarks that humans can do and AI cannot, and that job is getting harder. It’s not impossible yet. We have plans for ARC 4."
"There’s two ideologies within the field right now. One is you’re gonna be able to create an RL environment for anything and just hill climb that forever."
"Capitalism is running a really big hill climbing experiment right now. The labs find a gap, the RL companies fill it, and you go back and forth."
"On the flip side is, once the algorithm to AGI is found and you have an actual human learner, that algorithm will be able to meta learn anything."
"ARC Prize, we’re more interested in almost the romantic pursuit of that general learner."
"I have a feeling it will be a different branch of the tech tree than what we’re currently going down. The shape of what we have right now will be like 98% of the solution, but then it’s gonna be a repackaging in some format that will bring us there."
@arcprize
ARC Prize President @GregKamradt reveals how Astra invented its own programming language to score nearly 99% on ARC-AGI-3:
"At our ARC-AGI-3 launch in March, we had Sam Altman on stage to do a fireside with François. The moderator said, how long do you think it’ll take? And François’ answer was, well, it depends how much the labs make it a priority."
"We don’t have any concrete proof or evidence, but there’s an open question as to how much targeting was actually done on it. That’s the name of the game when you do these types of benchmarks."
"When Astra played this, it made its own domain-specific language on the fly. It created its own little code language for itself. That was something we hadn’t seen from models beforehand."
"Back in May, the models at that time thought they were playing Tetris. They thought they were playing Frogger. And then they tried to execute actions like they were playing those other games."
"Astra didn’t do that at all. Astra was much more aware about its novel environment."
@arcprize
GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI:
- Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness
- It surpasses human performance on 96% of ARC-AGI-3 levels
- It builds the most precise symbolic model of novel environments we've seen
Our analysis:
ARC Prize President @GregKamradt explains why Astra scoring 99% on ARC-AGI-3 doesn't mean AI has achieved AGI:
"Astra was very good at ARC-AGI-3 using one of OpenAI’s harnesses. It got something like 99%, so more or less it saturated the benchmark."
"The reason why is because ARC-AGI-3 games had a finite ceiling. There was one answer to go and do. You didn’t have to produce anything along the way."
"Each one of the games were simple. A human can do them in five to 10 minutes. We tested 500 humans on ARC-AGI-3, so we have human baselines."
"ARC-AGI-3 was a really good stepping stone in the journey of expanding the diversity of environments. But it was a stepping stone, not the final thing."
@arcprize
GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI:
- Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness
- It surpasses human performance on 96% of ARC-AGI-3 levels
- It builds the most precise symbolic model of novel environments we've seen
Our analysis:
SITUATION EXPLAINED: An early Anthropic employee says a co-founder got upset about office sushi night: "We’re not here to get rich and eat caviar."
• Ivan Vendrov joined Anthropic in October 2022, when it was about 40 people on 2 floors of an office in San Francisco’s Financial District
• His recruiter asked if he’d pledge to donate 2/3 of his equity, and it felt "a bit ambiguous" whether the offer depended on his answer, so he said yes
• At lunch, one employee said everyone was there because of AI x-risk, and a teammate replied that some just thought the tech was cool, which led to an awkward pause
• A senior colleague floated a new team to train dangerous bio, cyber, and nuclear models and show scary demos to governments
• When he said he’d rather work on beneficial deployments, his manager told him: "We are going to build AGI. And we’re going to do it safely."
• He quit after 8 months, before his equity cliff, when Anthropic was valued at around $4B
@theojaffee: "We’re not here to get rich. Oops. Well, now all the co-founders are gonna be worth probably about $50 billion at IPO."
SF Compute CEO @evanjconrad explains why Anthropic winning the AI race could spell disaster for GPU cloud providers:
"I don’t think the clouds are gonna blow up suddenly or AI is doomed. But it doesn’t matter what I think. It matters what a lender thinks."
"Their biggest customers are ones who themselves want to be the GPU cloud. The frontier labs, the hyperscalers, the people who have investment-grade credit, will simply try to disintermediate you if they have the opportunity to."
"What if Anthropic just wins, and the clouds can’t sell to Anthropic because Anthropic just wants to do this, and they’re stuck selling to these other competitors who might die."
"Is that true? I don’t think that’s necessarily true. But this is the scenario that people are worried about."
@sfcompute
