- Dreamer4 (https://danijar.com/project/dreamer4/) is a promising direction (by a frontier lab)
- Surreal. Almost like an optimistic science fiction film.
After all, self driving cars have some of the highest positive to negative externality ratio of any modern technology
- Amanda Askell, Sholto Douglas have somewhat of a fan following on twitter
- ChatGPT seems like a huge distraction for OpenAI if their goal is transformative AI
IMO: the largest value creation from AGI won’t come from building a better shopping or travel assistant. The real pot of gold is in workflow / labor automation but obviously they can’t admit that openly.
- The problem is paternalism and assuming the user is too dumb to take control their privacy preferences.
The compliance of the cookie banner regulation has measurable negative externalities - one estimate suggests a EUR 14B/year productivity hit in the EU
Most modern browsers allow you to disable all cookies if you like. You can always use incognito mode if you want to be selective about it.
In an ideal world, the EU could have simply educated their constituents about privacy controls available in their browser.
- They do have unreleased Olympiad Gold-winning models that are definitely better than GPT5.
TBD if that performance generalizes to other real world tasks.
- "Exoskeleton" was inspired by the more recent Dex-Op paper (https://dex-op.github.io/)
Calling UMI an "exoskeleton" might be a stretch but the principle is the same - humans use a kinematically matched instrumented end affector to collect data that can be trivially replayed on the robot.
- The exoskeletons are instrumented to match the kinematics and sensor suite of the actual robot gripper. You can trivially train a model on human collected gripper data and replay it on the robot.
- Kinda genius to scale exoskeleton data collection with UMI grippers when most labs are chasing "general" VLMs / VLAs by training on human demonstration videos.
Imo the latter will be very useful for semantic planning and reasoning, but only after manipulation is solved.
A ballpark cost estimate -
- $10 to $20 hourly wages for the data collectors
- $100,000 to $200,000 per day for 10,000 hours of data
- ~1,500 to 2,500 data collectors doing 4 to 6 hours daily
- $750K to $1.25M on hardware costs at $500 per gripper
Fully loaded cost between $4M to $8M for 270,000 hours of data.
Not bad considering the alternatives.
For example, teleoperation is way less efficient - it's 5x-6x slower than human demos, and 2x-3x more expensive per hour of operator time. But could become feasible after low-level and mid-level manipulation and task planning is solved.
- There is probably an equivalent of Amdahl's law for GDP - overall productivity will be bottlenecked by the least productive sectors.
Until AI becomes physically embodied, that would mean all high-mix, low-volume physical labor is likely going to become a lot more valuable in the mid-term.
- Ben’s original take about 1% of users being creators might end up being right eventually
Consider the Studio Ghibli phenomenon. It was fun to create and share photos of your loved ones in Ghibli aesthetics until that novelty wore off
Video models arguably have a lot more novelty to juice. But they will eventually get boring once you have explored the usually finite latent space of interesting content
- Math comparing new datacenter capacity to electric cars -
Projections estimate anywhere between 10GW to 30GW of US datacenter buildup over the next few years
1GW of continuous power can support uniform draw from ~2.6M Tesla Model 3s assuming 12,000 miles per year, 250Wh/mile.
So 26M on the lower end, 80M Model 3s on the upper end.
That's 10x-30x the cumulative number of Model 3s sold so far
And remember all datacenter draw is concentrated. It will disproportionately going to impact regions where they're being built.
We need new, clean power sources yesterday
- > consumers paying for electricity used by server farms
wait what? consumers are literally paying for server farms? this isn't a supply-demand gap?
- Also the pretrained LLM (the one trained to predict next token of raw text) is not the one that most people use
A lot of clever LLM post training seems to steer the model towards becoming excellent improv artists which can lead to “surprise” if prompted well
- The code isn’t the valuable part. They know all the most common workflows and failure modes, allowing them to create better environments for training agentic models
- > we should always promote low switching costs between model providers (by supporting independent companies), keeping incentives toward improving the models not ui/data/network lock-in
You’re underestimating the dollars at play here. With cursor routing all your tokens, they will become a foundation model play sooner than you may think
- Uh that doesn't look better. it has more texture but the composition is bad/incomplete
- Possibly. Seems like a mix of conspiracy-theory induced paranoia and right-wing influencers pushing a coordinated narrative.
Ironically, aerosol injection will probably benefit fossil fuel companies, with less pressure to meet aggressive emissions targets.
- the uproar over minor, localized cloud-seeding (which had nothing to do with the Texas floods) is probably a death knell for aerosol injection.
we are going to see countries going to war over unilateral solar radiation management efforts
I also suspect any future congestion and land-use problems will get better after an initial dip. Urban living becomes more desirable as city parking lots disappear.
If roads are used exclusively for self driving cars, this would probably improve traffic flow. Robot cars multiply current highway and city street capacity by coordination. They can smoothen traffic flow due to hard braking, and drive much closer to each other.