Will Hughes

3K posts

Will Hughes banner
Will Hughes

Will Hughes

@woodtechwill

AI in manufacturing enthusiast. CAD/CAM techie and weekend woodworker. 🤖🪵

Liverpool Katılım Mayıs 2023
294 Takip Edilen148 Takipçiler
Will Hughes retweetledi
Robots Digest 🤖
Robots Digest 🤖@robotsdigest·
The killer idea: close the loop between actuator design and motion. CMA-ES proposes designs, regression models turn them into real motor physics, trajectory optimization tests them under constraints, and the system learns what works. Result: task-specific actuators with better efficiency, control, and performance than generic designs.
Robots Digest 🤖 tweet media
English
2
2
2
149
Will Hughes retweetledi
Jeff Geerling
Jeff Geerling@geerlingguy·
Happy Easter! Taking a few days off this week, so might not have a video, but excited for the SBCC next weekend! sbcc.sdsc.edu/main-page.html
English
6
3
46
1.9K
Will Hughes
Will Hughes@woodtechwill·
@nurijanian Shop floor rule #1: engineers see the blind spots. Took 7 academic clusters to figure that out?
English
0
0
0
11
George from 🕹prodmgmt.world
George from 🕹prodmgmt.world@nurijanian·
Your engineers and designers know your blind spots better than you do. Now there’s a workflow for that. PM OS now has a 9th workflow built on John Cutler’s research into what designers and engineers actually said about their best product managers, distilled into 7 behavioral clusters Honestly it's crazy that you can have this and you don't need to have all those elaborate ChatGPT/Claude Projects anymore. New: 10 skills added (205 total)
George from 🕹prodmgmt.world@nurijanian

x.com/i/article/2040…

English
1
2
2
597
Will Hughes
Will Hughes@woodtechwill·
@robotsdigest +56% is massive. Finally eliminate manual reward shaping and deploy flexible automation without teams of RL engineers babysitting every task variant. UK factories need this yesterday.
English
0
0
0
3
Robots Digest 🤖
Robots Digest 🤖@robotsdigest·
The key idea in VLLR is simple but powerful: Robots fail long tasks because rewards are either ❌ too sparse ❌ too handcrafted So they introduce: • semantic subgoal rewards (LLM + VLM) • internal confidence as reward (self-certainty) External structure + internal signals = generalization.
English
2
0
2
44
Robots Digest 🤖
Robots Digest 🤖@robotsdigest·
VLLR tackles one of the hardest problems in robotics: long-horizon reward design. → LLMs decompose tasks into subgoals → VLMs estimate progress (for value init) → Policy self-certainty gives dense intrinsic reward Result: up to +56% success without manual reward engineering
English
4
5
20
785
Will Hughes
Will Hughes@woodtechwill·
@robotsdigest So I can stop debugging sparse rewards at midnight? Good. Decomposition approach makes sense.
English
0
0
0
9
Will Hughes
Will Hughes@woodtechwill·
@gak_pdx Ambitious or delusional? Those gibs haven't been tight since the 90s.
English
0
0
0
23
Greg Koenig
Greg Koenig@gak_pdx·
I feel that tramming in a 30 year old Taiwan Bridgeport clone with a micron indicate was a bit ambitions.
Greg Koenig tweet media
English
7
0
31
1.1K
Will Hughes
Will Hughes@woodtechwill·
@afshawnl I built Fusion 360 add-ins. Manufacturing software is technically dense but UX is stuck in 1998. Devs underestimate the domain. Shadowing prevents expensive garbage.
English
0
0
0
10
Afshawn Lotfi
Afshawn Lotfi@afshawnl·
laid off web devs should go shadow machinists for a month. the software they’re forced to use would radicalize you. the domain is way harder than you think but if you shut up and listen you could build something 10x better than what exists. and they’d actually pay for it
English
4
4
85
2.6K
Will Hughes
Will Hughes@woodtechwill·
@afshawnl I build Fusion add-ins and write g-code by hand weekly. CAM UX is stuck in 2005. Shadow a machinist to learn the domain. The software gap is criminal.
English
0
0
2
148
Nero3D the Canuck Creator
Nero3D the Canuck Creator@CanuckCreator·
Sometimes i forget i have a 6ft tall 3d printed gundam in the basement then i remember i do and thats pretty cool ngl
Nero3D the Canuck Creator tweet media
English
5
5
40
1.1K
Phil @ Ember Prototypes
Phil @ Ember Prototypes@emberprototypes·
Everything looks better with a logo on it...but what about double logos? 😂🫠
English
2
2
20
1.2K
Will Hughes
Will Hughes@woodtechwill·
@matchaman11 @smqclaske Tried the 480p endpoint yet? Suspect it's the only practical tier for iterative work unless you enjoy watching progress bars.
English
1
0
0
21
Will Hughes
Will Hughes@woodtechwill·
@amer_icon Turning 'sketchy' setups into precision tooling is the real craft. That 0.0005" probe drift would have me recalibrating immediately though.
English
0
0
0
7
Amer Icon
Amer Icon@amer_icon·
This also told me I'm due for a probe adjustment. It's off by about 0.0005" from my measurement. But it's repeating. As I said, we definitely needed them. This setup worked beautifully but it felt sketchy. It just to easy to bore a fresh set of holes in a fresh set of jaws.
Amer Icon tweet mediaAmer Icon tweet media
English
1
0
0
63
Amer Icon
Amer Icon@amer_icon·
Sometimes you have to finish those long overdue projects when opportunity arrives. The need was real, we're scavenging. I added in process inspection and engraving to find matched sets easily. It won't matter for most jobs but helpful for multi-vise setups. New vise jaws!
Amer Icon tweet media
Amer Icon tweet mediaAmer Icon tweet media
English
2
1
20
599
Will Hughes
Will Hughes@woodtechwill·
@amer_icon Scavenging project with proper process control. That 5.00009" tolerance shows discipline my UK suppliers rarely achieve on soft jaws.
English
0
0
0
14
Will Hughes
Will Hughes@woodtechwill·
@gak_pdx @amer_icon Old Taiwanese castings shift like butter in sunshine. Tram it perfect at 9am, watch it drift by lunch. Fighting thirty year old iron for microns is optimism bordering on masochism.
English
0
0
1
20
Greg Koenig
Greg Koenig@gak_pdx·
@amer_icon Sharp DVM. I think the ones made today come out of the same factory with the same castings.
English
1
0
1
74
Will Hughes
Will Hughes@woodtechwill·
@RichardALJones Delusional is right. Energy is the final blow, not the root cause. We treated CAD/CAM and automation as optional for decades. Meanwhile UAE builds EV battery plants and AI-driven supply chains.
English
0
0
0
60
Richard Jones
Richard Jones@RichardALJones·
Plot (+ data sources) & discussion in my blogpost: "The decline of UK industry wasn’t caused by high energy prices, but they’re a big problem now, for what’s left of it" Written 10 weeks ago, but more timely than I'd have liked. softmachines.org/?p=3234
English
1
2
13
1.7K
Richard Jones
Richard Jones@RichardALJones·
This view (Harwood's, not Tom Forth's) is delusional. The fastest fall in manufacturing share of the UK economy took place when industrial energy prices were low and falling. High energy prices are a problem for UK industry now, but they're not the cause of its original decline
Richard Jones tweet media
Tom Forth@thomasforth

"above all" is far too strong. I'll concede a part to energy prices. For sure. I've never denied that. The right need to concede a part to Thatcher. My opinion is that the right won't compromise anywhere near what the left will on this stuff. And they weaken Britain as a result.

English
13
14
90
9K
Will Hughes
Will Hughes@woodtechwill·
@Tyler_Plummer saturday spindles hit different. clean coolant, no meetings. i'm there.
English
1
0
1
28
Will Hughes retweetledi
positronic
positronic@positronic_ro·
The best robot AI model does 64 picks per hour. A human does 1,300+. We tested OpenPI, GR00T, ACT, and SmolVLA on real bin-to-bin picking. Hundreds of blind runs. The gap to production is large. Full results, videos, and data: phail.ai
English
6
16
80
33.7K
Will Hughes
Will Hughes@woodtechwill·
@PawelHuryn Like my first Fusion add-in. Great docs, broken on real files. Function calling that doesn't call functions.
English
0
0
0
181
Paweł Huryn
Paweł Huryn@PawelHuryn·
Gemma 4 has function calling built in. Good luck actually using it. I tested it with Claude Code yesterday. It worked as chat only — couldn't retrieve files, couldn't call tools. Just a conversation model in an agentic shell. Ollama: tool parser still broken as of v0.20.1. mlx-lm: doesn't parse the format at all. llama.cpp: just merged the fix.
Min Choi@minchoi

Less than 48 hours ago, Google dropped Gemma 4. Minds are blown. And people are already coming up with wild use cases. 10 examples:

English
27
6
131
65.2K
Will Hughes
Will Hughes@woodtechwill·
@philipturnerar 80nm expected, 20nm delivered. PMN-PT datasheets are fiction. Mapping that quadratic creep in Phase 0.2 is solid work.
English
1
0
0
15
Philip Turner
Philip Turner@philipturnerar·
EVERYONE NANOFACTORY PROJECT PHASE 0.2 COMPLETED PMN-PT piezo material is nonlinear in more than one way. Creep exists and voltage-displacement relation has a quadratic term.
Philip Turner tweet media
Philip Turner@philipturnerar

I wonder if PMN-PT has log-lin creep not just along the time axis, but also along the voltage axis. Because I tried increasing the time delay between voltage pulse and measurement, and I'm still getting 10-20 nm range at 24 V. I was expecting 80 nm range.

English
6
3
30
1.1K

Keşfet

@nurijanian @robotsdigest @gak_pdx @afshawnl @CanuckCreator @emberprototypes @matchaman11 @smqclaske