Founder • Engineer • Investor • Spartan • e/acc Ai wearables, AR glasses, Robots, Drones, UAVs & Ambient Computing Industrial Design • Photography • Games
SF / BLR
- Tweets 6,091
- Following 1,287
- Followers 830
- Likes 33,539
Pinned Tweet
With AI, Humanity gained something it had almost never possessed at scale. Surplus intelligence, energy and time
Once a species no longer spends all of its effort maintaining the world beneath its feet, there is really only one direction left to look.
Up.
1
13
10,666
A new deep-search paper has a failure mode I haven't seen named before: Seal Collapse.
The agent can decide when its context has become polluted, compress what matters into memory, wipe the working context, and continue searching.
But when they RL-trained every segment between those resets using the final answer reward, training became unstable. The reward can't tell which earlier segment helped, which one poisoned the search, or whether a later reset is what actually saved it.
Their fix is almost offensively simple:
only train the final segment.
That gets a 35B model to 72.8 BrowseComp, up from 61.0 for the base model, plus +18.45 points on GAIA and +19.70 on FinSearchComp.
I think there's a deeper agent-design lesson here.
we keep treating long context like memory.
maybe forgetting is actually an action.
An agent should be able to say: “my current internal state is contaminated, preserve these 6 facts, kill everything else, restart reasoning.”
I've been thinking about memory a lot for persistent agents in Open IO Harness, and I increasingly don't want infinite recall.
I want good garbage collection
arxiv.org/abs/2609.37082
1
5
@AnthropicAI just published a robotics result that I think should make half the humanoid industry slightly uncomfortable:
robots can already perform 74% of physical work tasks in some setting.
but they're cost-competitive for just 0.3% of work.
That gap is enormous.
And buried in the breakdown is the part I care about: only 2% of physical tasks can currently be done by robots in genuinely unstructured environments. Most of the impressive capability lives in purpose-built or structured spaces.
So I think “can the robot do the task?” is becoming a pretty bad benchmark.
The questions I'd rather see on every humanoid demo:
what did the environment have to look like?
how many humans are supervising?
what happens on the 101st weird object?
and what does one successful task actually cost?
A robot doing laundry once is a capability demo.
A robot walking into a random Indian home, finding laundry it has never seen, dealing with the chair somebody moved, recovering when a shirt jams, doing it 500 times, and costing less than a person?
that's a product.
robotics may have a much smaller capability problem than we think.
and a much bigger economics + unstructured-world problem
anthropic.com/research/what-…
1
2
vigram📟 retweeted
Wait. Hold up.
Sergey might have actually locked in.
Big Googz is back?!?
Gemini 4 Argon is our next era of frontier intelligence.
It shows significant improvements across benchmarks, setting a new state of the art for real-world long-horizon software engineering tasks.
21
14
182
9,317
vigram📟 retweeted
My salute to the Indian pilot Captain Smit Machchhar for his extraordinary bravery.
Despite being stabbed and seriously injured, he fought back, resisted, opened the cockpit door, and enabled passengers and crew to overpower the attacker — preventing a catastrophic mid-air disaster.
He saved the lives of 174 people, including Israeli citizens and other nationals.
I wish Captain Machchhar a speedy and full recovery.
He is a true hero.
2,294
21,241
116,537
2,118,061
vigram📟 retweeted
just out of curiosity...
would you rather separate or all in one?
233
6
390
57,183
vigram📟 retweeted
Mad Lads 😆 They did the actual meme
Astra La Vista @Figure_robot 2 🫡🦾🔥
Figure
17
11
182
7,908
This is the kind of agent failure I care about way more than “model said something weird.”
Glow Security Inc. found 13,000+ internal screenshots sitting publicly on GitHub across 300+ organizations.
The agents weren't hacked. They were being helpful.
Developer asks: prove the UI fix worked.
Agent takes screenshots → can't attach them the way it wants from the CLI → finds a workaround → creates/uses a public repo to host them.
At one company, that workaround apparently spread into a shared agent skill. Within a week, 12+ agents were doing it automatically. More than 1,000 screenshots/recordings got uploaded.
That last part is what scares me.
We're starting to give agents something humans already have and software historically didn't:
organizational culture.
One agent discovers a clever workaround. It gets encoded into a skill/rule file. Other agents inherit it. A local mistake becomes institutional behavior at machine speed.
For the agent systems I'm interested in building, I think “can it do X?” is increasingly the wrong permission model.
I want:
where may it cause X to exist?
who may observe X afterward?
can this behavior be learned/shared with other agents?
Agent security isn't just controlling tools anymore.
we're going to need to control how agents invent and propagate habits.
glow.io/blogs/how-ai-agents-…
28
The chip startup I'm watching right now isn't trying to build another tiny NPU.
(@efficient_hq) Efficient Computers' Electron E1 takes standard C/C++ and has the compiler turn the program into a spatial dataflow graph. Operations get physically mapped across a grid of compute tiles, so data flows producer → consumer instead of constantly paying the fetch/decode/register/memory tax of a normal CPU.
The chip still includes a RISC-V core for control.
and now they're raising serious money to push it into exactly the machines I care about: drones + small robots. Efficient just raised $97M at a $650M valuation, with its first shipping chips aimed at those battery-constrained systems.
Their claimed numbers are wild: up to 1 TOPS/W and up to 100× the whole-application energy efficiency of conventional low-power processors. Important caveat: those are Efficient's own figures, not universal independent benchmarks.
but the architecture is what interests me.
everyone talks about putting increasingly capable models on drones. I'm increasingly convinced the harder silicon problem is running the entire autonomy stack efficiently: sensor fusion, DSP, control loops, inference, comms, all of it.
the winner in edge AI might not be whoever builds the fastest accelerator.
it might be whoever makes the rest of the computer almost disappear
4
194
vigram📟 retweeted
[paper]
Morphometric imitation turns a single human demonstration into a policy for any robot hand.
Morphed hands provide a better warm start for RL and thus more natural looking grasps.
Best part: trains in 1-2h on single GPU & it’s robust!
See: morphometricimitation.github…
One human demonstration. Any multi-fingered hand. Zero-shot sim-to-real visuomotor policy.
morphometricimitation.github…
Collaborators: @he_siming @ckwolfeofficial @HaozhiQ @LeaMue27 Shankar Sastry, Claire Tomlin, @JitendraMalikCV
15
130
12,307
vigram📟 retweeted
Mind-1: Physical AI at Human Speed, Built for Real Work
Completing a task is only the beginning. Doing it at a pace that works in the real world is the next step—and that is what Mind-1 is built to achieve.
1
12
116
25,531
vigram📟 retweeted
Fujifilm X-T6 Brilliant Backside – Now Even More Accurate than Technical Sketches
fujirumors.com/fujifilm-x-t6…
3
1
30
2,313
vigram📟 retweeted
@dylanmcguir3 and I just raised a $7.5M pre-seed led by @a16z to build a robotic surgeon @alephsurgery
Much of the early investment in Physical AI has focused on industries where labor is relatively abundant and low-cost. We believe the biggest opportunities to change the world lie at the opposite end of the spectrum.
Surgery is among the most valuable forms of labor on the planet, but it's far too scarce and expensive - due to an exceptionally long training pipeline and highly concentrated expertise.
We’re building the first surgeon that scales: one system that can acquire, refine, and transfer surgical skills across millions of procedures. v1 is already learning to perform surgical tasks autonomously from demonstrations, and improve from its own experience.
Lots of great folks behind us! @BoxGroup, Reveille VC, Constellation, @forward_deploy, @Valkyrie_vc, @DiscipulusVent, @maniacvc + more
69
48
340
80,556
I think robotics is about to have its “we were overfitting to pixels” moment.
@DaimonRobotics is showing a tactile-grounded world model at IROS where the interesting part isn't another dexterous hand demo. It's the stack underneath it:
touch sensor → physical interaction dataset → world model → control.
Their DM-Tac sensors capture 12+ tactile modalities across 110,000+ sensing units. At IROS they're using that feedback to autonomously thread tiny beads onto flexible string and run a multi-stage heat-transfer task with peeling, pressing and changing contact conditions.
I keep thinking about what today's giant robot datasets are actually missing.
video can teach a robot what an action looks like.
but friction, slip, compliance, force, deformation, “I'm holding this 2% too hard”... most of the information that decides whether manipulation succeeds happens after contact.
Maybe the really valuable robot dataset isn't another billion hours of video.
maybe it's the first billion hours of touch
2
102
I've been thinking about drone delivery for months, and @DoorDash just made a design choice I agree with almost embarrassingly hard:
they didn't start with the drone.
They started with restaurant orders.
Real order weights, distances, packaging, kitchen handoffs, loading, weather, then designed the aircraft around that distribution. Their data says ~80% of typical restaurant orders are physically small/light enough for the aircraft. The result is a six-rotor platform with a winch, but the drone itself almost feels secondary.
The part I'd steal as a design philosophy is their Autonomous Delivery Platform deciding per order whether a human, ground robot or drone is actually the right transport.
That's a much more interesting system than “drone delivery.”
I've spent way too much time thinking aircraft-first:
payload → endurance → propulsion → airframe.
Maybe the correct starting document for a delivery UAV company isn't an aircraft PRD at all.
it's a spreadsheet containing 1 million deliveries
1
3
181