Ben Lin @benlatrobe

Interested in how ideas become real products | Engineering leader: JDM/ODM, NPI, embedded systems | AI-assisted hardware engineering | Husband. Father. Tampa Bay, FL Joined May 2007

Tweets

233
Followers

13
Following

250
Likes

492

Ben Lin @benlatrobe

a week ago

An agent saying "done" means its script ran. A 3D render means the geometry exists. Neither proves the design is ready to build. I treat execution, evidence, and approval as separate authorities, because hardware gets expensive every time one of them stands in for the rest.

0 0 0 1 0

View Details

Ben Lin @benlatrobe

a week ago

@0x50484244 Placement is where the design intent lives, so that's the step I'd never hand over blind. My rule: every critical placement gets a one-line 'why' — which requirement or constraint put it there. The LLM can move copper. It can't remember why.

0 0 0 5 0

View Details

Ben Lin @benlatrobe

a week ago

@dario_fresu Teach it to say "I don't know." The failure mode I worry about is a confident layout with no link back to why: which requirement drove that placement, what constraint set the stackup. Routing is solved. Traceability from intent to copper is the gap.

1 0 1 16 0

View Details

Ben Lin @benlatrobe

a week ago

@markfrancisio Sunlight is the bigger problem. Direct sun adds 20 to 30C on top of the chips' own heat. Give the 4G modem a solid thermal path into the case wall, it spikes hardest on transmit. ASA or PC, not PLA. Test in real sun at the worst angle, not on a bench.

0 0 0 6 0

View Details

Ben Lin @benlatrobe

a week ago

@yusan_lin @LiamFedus Which one is you?

0 0 0 39 0

View Details

Ben Lin @benlatrobe

a week ago

When two teams fight over engineers, headcount is rarely the real problem. It's usually fuzzy priorities, or a plan that assumed nothing goes wrong. Adding people to a confused plan just makes the confusion faster.

0 0 0 6 0

View Details

Ben Lin @benlatrobe

a week ago

A working prototype is a good start. Before a build, I want to know whether the materials are ready, the assembly steps are clear, and the test has a defined pass condition. Design freeze alone doesn’t answer those questions.

0 0 0 8 0

View Details

Ben Lin @benlatrobe

2 weeks ago

@hasegaw The scope still gets the final say. Keeping those measurements tied to the design revision makes the next iteration much easier.

0 0 0 8 0

View Details

Ben Lin @benlatrobe

4 weeks ago

@0xlangeai Code: HB9CB6

0 0 0 1 0

View Details

Ben Lin @benlatrobe

4 weeks ago

@berryxia HB9CB6

0 0 0 1 0

View Details

Ben Lin @benlatrobe

a month ago

@Peter05704721 generation is only the first step The harder part is building a workflow around requirements, constraints, deterministic checks, and physical validation so we know where the AI is wrong, not just where it looks right.

0 0 0 4 0

View Details

Ben Lin @benlatrobe

a month ago

@Peter05704721 很認同「怎麼知道它做對了」才是真正困難的部分。我自己在嘗試 AI-assisted hardware design 時，也越來越覺得 generation 只是第一步。真正重要的是把 requirements、constraints、validation 和實際測試串起來。 AI 不一定第一次就要做對，但我們需要知道它哪裡做錯了。

0 0 0 4 0

View Details

Ben Lin @benlatrobe

2 years ago

@karpathy How would you address the challenge of replacing hardware drivers with a neural net? Would love to hear more about the implementation details.