ran an agent, used a tool. bushu's dad. dev rel at @Opera. prev IoT cybersec @ Bitdefender. made a museum for Silicon Valley billboards: https://nitter.cf/t.co/Fq8xqmjIee
San Francisco
Joined April 2007
- Tweets4.4K
- Following1K
- Followers2.6K
- Likes26.6K
Pinned Tweet
the page I built for fun, 101ads.org, got a mention in @sfstandard today as the billboards on highway 101 are back in the general discourse
this takes me back to early childhood, to Salonta, a small town on the hungarian border. i haven't had it since.
Dan retweeted
my reachy responds ~2x faster now, with the same answers. 🤖
I got multi-token prediction for Gemma 4 E4B on a the jetson orin nx inside the Reachy Mini working - decode 26 to 54–65 tok/s, tool calls respond 0.7 s sooner, identical outputs, same tool-eval score. Fully on-device.
the southernmost billboard on 101ads org is NOT an AI ad, it's Poppy Bank: 101ads.org/b/219486d1/?s=102…
no. the early majority isn't into CLI. but you can imagine a fluid, tailored, UI/experience/website generated for you on demand in the near future
i don't get this take. don't we ideally want to be able to want to discern between human and machine output + appropriation?
"That AI engines ask users to decide which model to use is a sign we aren’t there. "
The company not obsessing about the benchmarks is slipping AI into all of the components users touch while others optimize a prompt-based CLI.
And they are doing it with compute that the users have paid them for.
If you’re counting Apple out because they aren’t spending hundreds of billions on infrastructure, you’re missing the point.
They are getting consumers to buy the connectivity and power for them.
Consumers care about experience, not benchmarks.
Outside of enthusiasts, cars have enough power and reliability for vast majority of users. They don’t care about the engine under the hood, just what the car does.
Parents care if they can wave their foot under the trunk to get it to open when there are arms are full. Ultrasonic? IR? Radar? Machine vision? Lidar? They don’t care, as long as it works reliably and is safe.
People don't buy horsepower. They buy what horsepower enables. And they certainly don't want to write a prompt to start the car.
That AI engines ask users to decide which model to use is a sign we aren’t there.
@gruber
Dan retweeted
Update on the Jetson-in-Reachy project: the brain had to get smaller. I benchmarked on the the 8 GB Orin Nano; the robot itself runs an Orin NX 16GB.
Gemma 4 on an 8 GB Jetson Orin Nano:
E2B
- 199 ms to first token on an 846-token prompt (stock: 485 ms)
- 42.7 tok/s decode (stock: ~27)
- 2.9 GB peak memory (stock: 6.7)
- MMLU within a point of unquantized
E4B (stock can’t even build it for the Nano)
- 636 ms to first token on an 846-token prompt
- 24.3 tok/s decode (llama.cpp: 14.1)
- ~4.7 GB peak memory
- MMLU-500 68.0%, same as unquantized
I put the engine files and the TensorRT-Edge LLM binary on Hugging Face, community build:
huggingface.co/mihaichiorean…
huggingface.co/mihaichiorean…
My project to put an Nvidia Jetson inside the Reachy Mini is progressing! One part left to re-print and I'll replace the base with a jetson brain. I'm excited! So is the "customer" in the armchair.
@ClementDelangue this feels timely given you all are joining forces with Nvidia. The reachy is great!!
next, @Ryanair introduces business class?
but can your agent fly without stopping for 8k miles over the ocean?
A juvenile bar-tailed godwit named "B6" set the world record for the longest non-stop migration by an animal by flying WITHOUT STOPPING 13,560 km (8,425 mi) from Alaska to Tasmania, Australia.
The journey took 11 days, in which time the bird lost half its bodyweight.
ALT https://www.usgs.gov/centers/alaska-science-center/news/juvenile-bar-tailed-godwit-b6-sets-world-record
I built a site to map silicon valley’s hwy 101 billboards and track the lore (80% are ai & tech).
one founder told me he first saw his board in the wild here.
now each ad has its own link, easy to share, like this OpenAI @steipete feature: 101ads.org/b/c8496954/
lots of money to be made building the plumbing for this. accurate timeline
Delta, Chase and Amazon are blocking AI agents.
The war on agents has begun, and here’s how it probably plays out:
Stage 1: Companies block agents because agents skip the ads and upsells.
Stage 2: Agents learn to act like humans, and it becomes an arms race.
Stage 3: The market splits. Companies that make money on sales let agents in, companies that make money on data charge for access, and the giants build their own agents.
Stage 4: A trusted middle layer for agent identity and payments becomes the standard way in.
Stage 5: Companies that welcome agents start winning customers from the ones that block them.
Stage 6: The agent blockers come back and let agents in on their own terms.
Stage 7: Everyone starts competing for agents with better prices, faster checkout and SEO built for agents.
Stage 8: The ads come back, and companies pay to be your agent’s first recommendation.
Stage 9: The agent war ends, and selling to agents becomes as normal as having a website.
