Harnesses. Workflows. Apps. Story. "But all things excellent are as difficult as they are rare." —Spinoza
Los Angeles
Joined March 2009
- Tweets13.3K
- Following142
- Followers1.6K
- Likes73
Pinned Tweet
"The hottest new programming language is English."
Karpathy said that three years ago. His AI Startup School keynote shows what it means: Software 1.0 was code. 2.0 was weights. 3.0 is plain English — and prompt engineering is real engineering.
🔗 below
2
5
125
Michael D. Olmos retweeted
here's how i shipped 2,500 PRs last month to production
this was originally supposed to be for Cursor Compile in London. i couldn't make it since i was livestreaming for Grok @Bot Galaxy so i'm making it available for free here on X! watch it on 2x speed, i talk slowly
ChatGPT now has a place to put the work, not just the chat. OpenAI calls it Space.
The demo: ask ChatGPT to turn a long thread into a page. Priorities, decisions made, what's still open. You edit it directly or ask ChatGPT to change it.
Then you share it. A teammate opens the page and updates it in real time. Pages is the new format, plans, drafts, dashboards and visuals in one spot.
Your chat stays private, and so do your sources. Your teammate sees the page, plus you and ChatGPT editing it. Not your conversation.
Last move: tell ChatGPT to keep the script current with Slack and Google Drive. (demo, so we'll see)
my read: the private chat, shared page split is what makes this workable for a team. i'd run one live project through it before trusting it with more.
the sharing part, around the 1:00 mark, is the part.
youtube.com/watch?v=G-9Y2xNi…
55
Quadrillions of transistors. That's the scale NVIDIA says chip makers are now designing and running.
Their 2:19 video walks the whole pipeline: simulation, digital twins, agents, robots, vision inspection.
Two numbers worth pulling out. Chamber simulations run up to 35 times faster. And NVIDIA Cosmos, a vision language model, hit over 96% accuracy classifying wafer map defects (their numbers, not mine).
The digital twin piece is the real idea - test the change in a shared virtual copy of the equipment and the fab before it touches the real one.
my read: it's a vendor pitch, so no benchmarks or methods. but simulate-first is a habit you can steal for any system where a bad change is expensive. start with the digital twin section.
youtube.com/watch?v=rBN0PDfJ…
56
Nvidia's own line on its new laptop chip: "We're entering an era where it's not about having a giant lab train a model. We're entering an era where you can train a model."
That's the pitch for RTX Spark.
Blackwell GPU plus a power-efficient CPU, with 128GB of memory. The most ever for an RTX GPU.
It's one unified memory system, not just a laptop with a GPU bolted on. That's what makes local models workable on a machine you carry.
The demo: a local model, 3D rendering and a game at over 100 frames per second, all multitasked. (their demo, not mine)
my read: unified memory is the spec to watch, not the frame rate. if local training on a thin laptop actually holds up, that's the part worth sitting with.
youtube.com/watch?v=vJvFta7e…
78
"Woohoo! Airmail straight in the hole."
That's a humanoid robot, sinking cornhole shots. DeepMind showed Gemini 4 one video of a person throwing a bean bag, then let it practice in a simulator.
It watched the video of each failed throw, figured out what went wrong, and tried again.
The four steps, per the robot's own recap: map the human arm swing onto its joints, measure the board with its two eye cameras, practice in simulation (wider stance, so the hand clears the hip), then adjust the release by a few cm on the real robot.
Seven shots straight in the hole.
My read: the loop is the story, not the cornhole. watch, fail, look at the footage, fix it - that's how I'd want any tool I use to learn.
the last 20 seconds are the part.
youtube.com/watch?v=9PICEmtf…
32
"I've spun up more changes into Chromium in the last two weeks than I would typically do in a matter of months."
That's an engineer with early Gemini 4 access, per a Google DeepMind video.
Chromium is about 40 million lines of code and two decades old. Earlier models kept going off the rails on it.
Now: no skills, no harness, just a memory file. It spins up 20 changes at a time.
Every change gets a couple hundred thousand tests, two human reviewers, then a couple million more tests on new and old devices. Twenty at a time pass.
A typical engineer puts out one change a day. 20 changes = 20 days of work.
my read: the test-and-review gauntlet is what makes this believable, more than any model claim.
worth the 2:35.
youtube.com/watch?v=LVEfFyij…
55
Lex Fridman just dropped a 90-minute trip through Taiwan with Jensen Huang: street food to the foundry business:
0:19 - the famous oyster pancake sets the scene
15:26 - the case for reindustrializing the United States
44:35 - where export controls come into the conversation
1:19:41 - how the foundry business enters the story
1:23:40 - why TSMC got picked over UMC
For anyone who wants the chip supply chain told over a lot of food.
my read: skip ahead to the foundry stretch, that's the part worth sitting with.
youtube.com/watch?v=FGCP2F_7…
81
"It's 2026 and you don't want to build your own agent platform."
That's the pitch from Databricks for Agent Bricks.
Their diagnosis: teams are wrestling Kubernetes, which was built for web services, and maintaining databases instead of working on agent context.
What's in the box:
- hosting with automatic recovery
- one API for switching between models
- sandboxes that launch in seconds
- managed agent memory
- ready-made MCP servers for Genie, search, Slack, GitHub
They claim scratch to a working agent in under 5 minutes (their number, not mine).
my read: the single model API is what I'd test first. the rest only matters if it holds up past the demo.
getting-started section is the part, and the whole video is under 3 minutes.
youtube.com/watch?v=OpMMcxcR…
34
Matthew Berman on Meta's Muse Wonder: "I did not expect Meta to have a Frontier model."
The benchmarks lean his way. On the Artificial Analysis intelligence index, Muse Wonderful ties Sonnet for second - 56, with Opus 5.5 at 58.
Muse Wonder on the API runs $4 per million input, $20 per million output. Wonderful is only available with a Muse Code subscription.
The hands-on part was rougher. His "hyperrealistic" watermelon render came out blurry. He kept hitting 429 errors on the $50 plan, then ran out of quota again on the $200 one.
my read: strong scorecard, shaky access. i'd wait for the quota mess to settle before building on it.
the benchmark walkthrough is the part to watch.
youtube.com/watch?v=rpLrkJ5N…
77
Code review is the new bottleneck. That's OpenAI's own framing, in a 2-minute Codex demo.
The pitch: one inbox for every GitHub pull request. Assigned to you, authored by you, priorities. Pin the one that matters.
Cloud review takes a first pass before you open it. Its comments land next to the change, so no tool-switching.
Then the part I'd actually use: each PR gets a private chat. Ask for a summary, or the risks across the whole project. Your team's review guidelines and design docs come along as context.
You can even have it explain the change as a comic book (we'll see).
GitHub Code Review is generally available, GitLab is in preview.
my read: the inbox is the easy win. the private chat is where I'd test whether it catches what a tired reviewer misses.
youtube.com/watch?v=IZXv6a78…
40
a16z just dropped a 32-minute conversation on recursive self-improvement, from what RSI means to how you'd measure it:
2:39 - what people actually mean when they say RSI
5:35 - the numerator and denominator of improvement
11:26 - a "bleederboard" for what experiments cost
16:55 - RSI at every level of the stack, not just weights
28:21 - Amdahl's law, and why wet-lab work caps the speedup
For anyone who's read an RSI announcement and wondered what it actually proves.
my read: the cost side is the part worth sitting with.
youtube.com/watch?v=3Q1mdn5Z…
23
RTX Spark let her finish her grandfather's chess set.
He hand-carved the pieces. He passed before the set was complete.
She took an image of one piece, turned it into a 3D mesh in ComfyUI, and kept Blender open on the same machine to design the missing ones. Then straight to her 3D printer.
The prints go to family as gifts.
my read: this is the kind of local-AI workflow worth copying - scan, mesh, design, print, one box, no handoffs.
youtube.com/watch?v=nM21x3FF…
1
98
"We started this with about two thirds of our Meta ad spend. Today, Genie and our agents are optimizing 100% of it."
That's HelloFresh, in a 97-second video from Databricks.
The before: data silos everywhere, and weeks until a decision on how to spend money.
The after:
A week of reconciliation is now under an hour on a Monday morning.
Ads get tested four times faster.
A seven-person launch process happens in one click.
Maintenance work that ate 40 to 50 hours a week now takes one or two (their numbers, not mine).
my read: the interesting part isn't the model, it's that they got the data access sorted first. the video credits Unity Catalog for that, and the rest follows from it.
youtube.com/watch?v=mF82aK4e…
47
Matthew Berman just dropped a 17-minute walkthrough on AI decompiling and rebuilding video games:
0:42 - source code vs machine code
3:20 - why guess-and-check is a fit for AI
5:02 - his own Super Mario World rebuild
7:09 - how a mod connects two running games
13:40 - recursive self-improvement, his "ultimate version of a loop"
For anyone curious what a long-running AI loop looks like in practice.
my read: the loop is the part to steal. if the answer can be checked automatically, it can be handed off.
the 3:20 section is enough to get it.
youtube.com/watch?v=cX2kD2yQ…
54
150+ ambassadors. 80+ countries. 300+ events.
The Codex Ambassador Program started with a plan for 15 to 20 people.
My favorite story in here: the Kochi ambassadors in India ran a whole Codex event on a running metro. Packed car, everyone coding together. Now it's starting a trend.
Then Build Week, tied to the GPT-5.6 release - 60+ events inside a week, popping up in Tokyo, India, London, SF.
Their one requirement for ambassadors: curious. You don't need to know everything about Codex yet.
my read: local, in-person scenes around a tool is how it actually spreads - and a coding event on a metro beats any launch webinar.
youtube.com/watch?v=enPbh_A5…
1
61
a16z just dropped a 53-minute talk with the former Intel CEO on building hardware, from his early chip-design years to data center energy:
1:51 - how his Intel career started
10:48 - why nothing is just a chip anymore
24:40 - memory as the AI story
35:02 - copper vs optical
43:26 - the energy capacity problem
49:11 - the next abstraction
for anyone trying to see where AI's physical bottlenecks actually sit.
my read: the copper vs optical stretch is the part worth your time.
youtube.com/watch?v=1Q_7yU7F…
1
101
dots are always-on agents that take work off your plate. now you can set one up from your phone.
Open the ChatGPT app, tap "Meet dot" in the top left, and it builds a thread and sets itself up. It may ask for calendar and email access. The presenter recommends connecting both - more context, more it can take off your plate.
You can name it and give it an avatar too. The demo dot is called Blossie.
Then the fun part. He calls Blossie and asks for an iOS widget prototype. It kicks off a Codex thread, checks Slack for who else is on the widget, and drafts a reply for him to approve.
my read: the approve-before-it-sends step is what makes me trust this. an agent that drafts and waits is useful. one that just fires isn't.
The call-and-delegate demo is the part to watch.
youtube.com/watch?v=p5KxBTUn…
110
Peter Diamandis just dropped a nearly three-hour episode, from chip speed to Nobel season:
0:30 - the "World War II moment" framing
9:54 - Cerebras and 1,200 tokens per second
38:01 - quadratic turning subquadratic
1:14:25 - Claude as a new kind of entity
1:33:14 - thousands of specialized AI agents
2:16:49 - Nobel Prize season
my read: nobody needs the whole thing. start at the Claude segment and treat the rest as a menu.
youtube.com/watch?v=LRdb8UmP…
118
Matthew Berman just dropped a 25-minute walkthrough of 12 Grok Bot use cases, from chores you hate to a bot that handles them:
0:46 - have it negotiate your bills down
2:38 - connect it to your Tesla
4:08 - a weekly bot that hunts stale files on your hard drive
8:29 - a family bot for school and sports emails
13:48 - a meeting summary bot that chases your action items
20:42 - compare electricity plans against your own usage
He reads out the prompts too.
my read: skip the Tesla one, start with the email and meeting bots - that's where the hours are.
youtube.com/watch?v=BsNHg6cN…
31
Microsoft just dropped an hour-long Windows event, with Satya Nadella and Jensen Huang on stage: from local models to agents running on your PC:
0:09 - the hybrid intelligence pitch
4:02 - a coding agent that picks a local model on its own
12:35 - Autopilot demo: a file clean-up that ends in an accountant's hands
32:41 - why Nvidia was founded
53:13 - why agents need an identity separate from the user
For anyone deciding whether agents belong on their own machine.
my read: if you only have five minutes, start at 12:35. that demo is the pitch.
youtube.com/watch?v=RE_AsVyo…
1
1
83