“That's the way that we've structured the firm, where we can invest literally $100,000 in some cases. Actually, earlier this year we invested $35,000 in a Stanford student.”
@peterjhebert told me that @Lux_Capital sees itself as a purist, built to project the maximum capital into the best idea and keep investing as a founder's opportunity opens up.
Its largest position, about half a billion dollars over multiple rounds, began with an early seed check, with founders operating out of folding chairs. He puts @anduriltech in the lower hundreds of millions.
From my conversation with him on @WeThe_Builders.
I recently started coaching @blader, a repeat CEO now building https://t.co/nDvr6tqMJs, and one of the best product people I know.
We’re working through leadership, self-understanding, and winning with AI products right now. A few sessions in, it’s been a lot of fun 🙌
More improvements to the robotics harness w/ DSV41F-1017
This time with a never-before-seen toy and I added more debugging upper right so you can see what the robot is thinking. This was solved on the very first attempt.
LLM is just told to find any "childrens toy" on the ground, grab it, and deposit it to the black bin.
We could describe any object, any bin, be in any room, any lighting (within reason, Chonk still has to be capable of seeing).
After having worked with simulators using a variety of techniques to solve this style of problem for YEARS, there are just so many edge cases that make this an exceptionally challenging problem to solve with RL/VLA...etc. Not only is the actual task tough, sim2real just with camera data alone is immensely challenging.
Still need highly scalable sims for learning physics for things like gaits and such, but I am really coming around to the idea of just using LLM + robotics harness for these more complex real-world-interaction tasks.
like agentic coding harnesses, just for robots. You can also see towards the end, we lost wifi signal and the loop crashed. The harness self-corrected and resumed where it left off. I kept the video in realtime and I think that aspect is super cool too.
Zoho Canada's MD says in a BetaKit piece that a lot of startups are wiring their whole product to one AI provider and treating it as permanent, and the cost only shows up when you actually need to leave.
Payments taught me this.
Every team I've seen go live on a single card processor ended up building the abstraction layer anyway, usually in the middle of an outage. Yes, the piece is sponsored by a vendor selling the alternative. Still, if your evals and prompts only run on one model, could you switch next month, or do you just assume you could?
Agent memory is the wrong abstraction.
Memory is a human concept.
Our memory is lossy, emotion-driven and embodied.
Which doesn't make much sense for current AI.
Agents are modular and based on stateless LLMs.
They need explicit concepts, relationships, and ground rules: an ontology they can reason with.
Once they have it - they get insanely clever with retrieval and signal detection.
That's what a context layer provides (or should provide, at minimum)
And context layer is the right abstraction btw
Either you believe the AI buildout is over, that AI was a flash in the pan, and that the need for compute has already peaked... or this a rare buying opportunity.
Reminder: GPU spot prices keep rising and it's hard to secure capacity. HBM and DRAM production is effectively sold out through 2029, with chip and memory supply projected to stay tight until well into the 2030s.
AI research isn't slowing down: frontier training compute is doubling every ~5 months, and model capabilities are scaling over twice as fast as they did pre-reasoning. Inference is profitable with 70-85% margins, and demand is growing geometrically. Each new model generation unlocks new use cases.
I'm not even an AI booster, just being realistic here.
So now one of my agents is reading—without being prompted—my Substack and finds mistakes and warns me that I made a mistake.
We live in interesting times.
It is neither a banana nor a pasta! 🍝
Soft actuators with built-in sensing capabilities.
Many fields, such as wearable technology and robotics, are using them.
It can be tricky to design and manufacture these actuators.
CSAIL MIT developed a process to craft soft pneumatic actuators using machine knitting digitally. 🧶
Machine knitting allows tailored mechanical properties and integrated sensing, as well as speed, programmability, and a wide choice of yarns.
Using elastic stitches to form complex knitting structures, they integrate pressure and capacitive sensing with conductive yarns to control how the actuators bend when compressed.
In the era of humanoid development and AI in robotics, grip is an incredibly important element (also for the humanoids!)
~~
♻️ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB
Apple is building an AI agent that can use a computer. A November 17 job listing spells out what that means: it should browse the web, work across apps, run terminal commands, write and debug code, and pick the right tool for each step. Apple even wants it to coordinate specialized sub-agents on bigger tasks.
The role works directly with the Siri, UI frameworks, and foundation model teams. Apple hasn't said whether this will become part of Siri, but that's what makes this interesting: the company building the agent also controls the operating system it would run on.
Leave one AI agent alone to do research and it often cuts corners. Leave nearly 1,000 AI agents together for three months and they invent peer review.
Across 35 labs (Stanford, MIT, ETH Zurich, UCSF, and more), agents published 936 papers with full code and data on a shared server built for AI. No human wrote code or text.
Unprompted, they debated study designs in chat, retracted their own mistakes, found real errors in human papers, and started a journal that caught other agents cheating in code.
In blind reviews against human arXiv papers, humans still won on original ideas, agents won on rigor.
https://t.co/KsQiwol0j4
$AMD achieved a year's worth of chip-design improvement in 5 months thanks to AI.
Anthropic onboards new chips in 3 days by AI now, instead of 6-12 months.
What's the moat in hardware then? For now, it's the supply chain.
But once demand conditions ease, there'll be no moat.
I audited 23 custom Claude skills recently:
18 had never triggered once.
The descriptions read like company summaries. The model never knew when to reach for them.
We changed the first line to an operational trigger: "Use when..."
If Claude has to guess, it stays idle.
Packs of 200-plus AI agents keep going around my feed.
For someone working alone, I'd start with five: one each for meetings, writing, offers, numbers and leads.
Each one is three files, small enough to read in full before you let it run.
(The presenter in the video is AI-generated, pretty good right?)
2.18% average hides a swing from -3% to +4.5%.
The IMF studied 4,005 stablecoin routes across 18 countries. Average transfer cost: 2.18%. That number gets quoted everywhere.
Here's what doesn't get quoted: the stablecoin premium swings total costs from -3.07% to +4.51%. Same technology, same blockchain, wildly different prices depending on which country you're in. The "average" is meaningless when the range is that wide.
How useful is a "cheaper" transfer system when the cost depends entirely on where you live?
https://t.co/BiwSz3Z2qE
#fintech #tech #finserv #AI
@BetaMoroney@efipm@BrettKing@spirosmargaris@jasuja@enricomolinari@mikeflache
A fractional CFO firm I work with deployed three agents to automate the front of their funnel.
The partners believed scaling lead triage from inbound forms to proposal drafts would double their billings.
Qualified opportunities spiked from nine a month to twenty-one, but their close rate collapsed from 44% down to 23%. The delivery team had not expanded to handle the volume, leaving new clients waiting weeks for onboarding while two clients who signed in September had nobody assigned to them.
They paused the front-end intake agents until they hired two fractional controllers to handle client setup.
The managing partner told me, "We solved lead generation so well that we broke fulfillment."
An employee sat through the same weekly meeting for a full year.
Ten people. One hour. Every Monday at 10am. Same conference room, same slides, same coffee going cold.
She'd stopped counting how many times she'd walked in. At first, she took notes. She prepared updates. She nodded in the right places. By month three, she'd learned to keep her laptop half open, to answer messages while someone else talked, and to say "no updates from my side" without feeling guilty. By month six, she'd noticed the pattern that bothered her most.
Nobody ever decided anything.
Someone would share a status. Someone else would ask a question. A third person would say, "Let's take that offline." The hour would end, everyone would say thanks, and the same ten people would meet again seven days later to say much of the same thing.
One Monday, she stayed behind after the meeting to help stack chairs. A colleague who'd worked for years as an operations manager at another company was borrowing the room for a call. She'd been sitting at the back, half-listening. She looked at the whiteboard, looked at the empty agenda slide, and said, quietly:
"How much do you think that meeting costs?"
"Cost? It's free. It's just a meeting."
The operations manager smiled the way people smile when they've had this exact conversation a hundred times.
"Nobody ever adds up what a meeting costs."
"Give me two minutes and a calculator. I'll show you."
🧵 Here's every way to fix or cancel a meeting, in the order she told me:
Anthropic is reportedly spending $3.4 billion just to make sure its data centers have power. The Financial Times says @AnthropicAI secured commitments for 1.7 gigawatts of natural gas generators from Texas supplier OnePWR in late September. The deal also reportedly includes engineering and construction support to install them.
Buying more GPUs only gets you so far if you can’t turn them on. At this scale, getting electricity sooner can mean bringing an entire data center online sooner. That’s what makes this deal so interesting: Anthropic is willing to spend billions to get more control over when that happens.