AI @anduriltech, Christian, optimist
Costa Mesa, CA
Joined November 2023
- Tweets1.3K
- Following988
- Followers403
- Likes4.2K
Jev/system-one models actually seem like a very promising solution for the monitor ability costing a lot of compute problem
Jack ๐บ๐ธ retweeted
Cal Poly keeps winning
Bradley Kern graduated from @CalPoly with a mechanical engineering degree. Before starting his master's this fall, he spent the summer at @freeform_future .
His assignment: an automated system to destroy Skyfall's massive air filters without manual handling or downtime. Bradley designed it from scratch: a 900kg carbon steel weldment, rated for 150,000 lbs of hydraulic force, that takes a 52-inch filter down to 8.
"I've done more here in a month than I've done in twelve weeks at NASA."
That's the kind of ownership we hand our interns at Freeform. Critical projects that outlast the internship itself.
Apply today. freeform.co/careers
#MechanicalEngineering #Hiring
Shovel seller declares America wins the gold rush by making U.S. shovels the global standard, arguing โitโs about which tools the world runs on.โ
Venture capitalist: I donโt think any of you are comprehending how successful my investment is going to be
Your favorite hard/deeptech company is an early RL environments company. I'm sorry if it pains you to hear this
Jack ๐บ๐ธ retweeted
humans, a bipedal primate notable for early contributions to matrix multiplication
Nearly every take here is wrong. Nike is failing for a much simpler reason
How to destroy a beloved brand.
Step 1: Put a soulless, data-driven consultant who has never even seen a basketball court or a running track up close in charge.
Step 2: He goes direct! And increase profits while destroying the retail distribution system and betraying a multitude of 20-year-long partnerships.
Step 3: Motivated and innovative competitors like On Running and Hoka rush to fill the distribution gap Nike left by abandoning its trusted partners.
Step 4: Not satisfied, he fires all the stupid creative people who can't prove any of their ideas work with hard data. What could they possibly know?
Step 5: Celebrate when milking the brand for every last cent works for about 18 months, even though the data quickly show a disaster is looming.
Step 6: OMG? Wut?
Step 7: Fire the old CEO and bring in a new one who's been with the brand since day one. He tries his best, but just too much damage was done.
Step 8: Get delisted... Go down in history as the greatest example of data-driven idiots destroying one of the most successful creative brands of all time.
The end.
Evidently I was wrong about this. OpenAI was training Astra with budgets in mind at the time in an effort to teach models about their own trajectory efficiency. You will also notice the shift from price per token to price per task. Good direction!
There seems to be a tradeoff between agentic capabilities and teaching capabilities, there are not good benchmarks for the latter yet
In case itโs not obvious already the transformer is one of the most important algorithms ever created and itโs not going anywhere. Worth trying to understand inside and out.
1. Robot data isn't the only way models can learn to control robots. Coding, math, computer-use all contain information about the physical world.
I bet that generalist LLMs will work better than robot-specific models for this reason
2. The output format (eg. joint angles, code-as-policy, tool calls as in Gemini ER) is up to debate. But why not just let an LLM choose?
