Thomas Wood
@advanced-eschatonics.com
AI scientist, roboticist, farmer, and political economist. Governments structure markets. IP is theft. @phytomech.com is my alt.
https://advanced-eschatonics.com
https://advanced-eschatonics.com
Disagree
anyway I do basically think nuclear is a stupid option for decarbonizing the US relative to solar and storage. those are reasonable and safe technologies that will still go on working as normal even if the government collapses or whatever
November 17, 2025 at 2:36 PM
Disagree
![timeout 20 curl -s -X POST http://localhost:7070/session/ses-a914983a-20a9-47d1-a002-d75463951585/message \
-H 'Content-Type: application/json' \
-d '{"agent": "build", "parts": [{"type": "text", "text": "What tools do you have? List all of them."}]}' | jq '.parts[-1].text' -r](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:6amf2dzllh6lvnsqxsr4nf6e/bafkreia45yzsspp5qowoxtkgd34zf5j42vlt56axizy575frft4whjffze@jpeg)
Reposted by Thomas Wood
Coming around to this viewpoint tbh
Hot take: we already have AGI, its just that people overestimate general intelligence
Let's recognize that peoples timelines for AGI have moved back. Those that held the shortest timelines should say "I was a little wrong" and ask why they were wrong.
November 16, 2025 at 10:04 PM
Coming around to this viewpoint tbh
Reposted by Thomas Wood
Thank you for saving us from communism sir
BESSENT: “THE $2,000 DIVIDEND COULD COME IN LOTS OF FORMS… IT COULD BE JUST THE TAX DECREASES. IT COULD BE NO TAX ON TIPS.”
November 17, 2025 at 2:22 AM
Thank you for saving us from communism sir
vibes and essays
BESSENT: “THE $2,000 DIVIDEND COULD COME IN LOTS OF FORMS… IT COULD BE JUST THE TAX DECREASES. IT COULD BE NO TAX ON TIPS.”
November 17, 2025 at 2:51 AM
vibes and essays
Reposted by Thomas Wood
BESSENT: REAL $2000 DIVIDEND WAS THE FRIENDS WE MADE ALONG THE WAY
BESSENT: “THE $2,000 DIVIDEND COULD COME IN LOTS OF FORMS… IT COULD BE JUST THE TAX DECREASES. IT COULD BE NO TAX ON TIPS.”
November 17, 2025 at 12:47 AM
BESSENT: REAL $2000 DIVIDEND WAS THE FRIENDS WE MADE ALONG THE WAY
## What is HOTL?
Human-Out-of-The-Loop programming:
1. Spec your project with the architect
2. Walk away (cook gumbo, hit the gym, watch a movie)
3. Come back to working software
4. Review the markdown exports
5. Ship it
Slow is smooth. Smooth is fast. Built for local inference.
Human-Out-of-The-Loop programming:
1. Spec your project with the architect
2. Walk away (cook gumbo, hit the gym, watch a movie)
3. Come back to working software
4. Review the markdown exports
5. Ship it
Slow is smooth. Smooth is fast. Built for local inference.
November 16, 2025 at 11:56 PM
## What is HOTL?
Human-Out-of-The-Loop programming:
1. Spec your project with the architect
2. Walk away (cook gumbo, hit the gym, watch a movie)
3. Come back to working software
4. Review the markdown exports
5. Ship it
Slow is smooth. Smooth is fast. Built for local inference.
Human-Out-of-The-Loop programming:
1. Spec your project with the architect
2. Walk away (cook gumbo, hit the gym, watch a movie)
3. Come back to working software
4. Review the markdown exports
5. Ship it
Slow is smooth. Smooth is fast. Built for local inference.
# Kill a runaway crow
murder kill frontend
murder kill frontend
November 16, 2025 at 11:52 PM
# Kill a runaway crow
murder kill frontend
murder kill frontend
Reposted by Thomas Wood
theres also ruka: ruka-hand.github.io
RUKA: Rethinking the Design of Humanoid Hands with Learning
A tendon-driven humanoid hand that is compact, affordable, and capable of dexterous manipulation.
ruka-hand.github.io
November 16, 2025 at 9:29 PM
theres also ruka: ruka-hand.github.io
Reposted by Thomas Wood
Yes, RUKA also seems to be great, unfortunately it's bit more expensive to build (still nothing in robotics world, but when talking abt hobby side project, AmazingHand seems more reasonable). Also recently came across aero, but their license is not straightforward oss: github.com/TetherIA/aer...
GitHub - TetherIA/aero-hand-open: https://shop.tetheria.ai/
https://shop.tetheria.ai/. Contribute to TetherIA/aero-hand-open development by creating an account on GitHub.
github.com
November 16, 2025 at 9:38 PM
Yes, RUKA also seems to be great, unfortunately it's bit more expensive to build (still nothing in robotics world, but when talking abt hobby side project, AmazingHand seems more reasonable). Also recently came across aero, but their license is not straightforward oss: github.com/TetherIA/aer...
Reposted by Thomas Wood
After I finish my current side-quest I'm planning to test out github.com/pollen-robot...
i think it has a potential
i think it has a potential
GitHub - pollen-robotics/AmazingHand: Code and model to control the AH!
Code and model to control the AH! Contribute to pollen-robotics/AmazingHand development by creating an account on GitHub.
github.com
November 16, 2025 at 9:17 PM
After I finish my current side-quest I'm planning to test out github.com/pollen-robot...
i think it has a potential
i think it has a potential
Reposted by Thomas Wood
Unitree H2 base model is < $30,000 USD. Incredibly cheap. Most people will buy the H2 EDU for I assume $50-60k but still
November 16, 2025 at 7:30 PM
Unitree H2 base model is < $30,000 USD. Incredibly cheap. Most people will buy the H2 EDU for I assume $50-60k but still
😃
November 16, 2025 at 8:09 PM
😃
Reposted by Thomas Wood
a majority of US adults can't reliably answer this question
but GPT can
but GPT can
November 16, 2025 at 7:43 PM
a majority of US adults can't reliably answer this question
but GPT can
but GPT can
Reposted by Thomas Wood
the wild part — this probably doesn’t have anything to do with AI or LLM architecture
e.g. the gospels of Matthew, Mark, Luke & John are all rephrasings of the same stories
Rabbinic teaching involved repetitive rephrasing
oral traditions in general are less about precise recitation
e.g. the gospels of Matthew, Mark, Luke & John are all rephrasings of the same stories
Rabbinic teaching involved repetitive rephrasing
oral traditions in general are less about precise recitation
Physics of Language Models: Part 3.1
If you show a fact to an LLM in pre-training once, it’ll memorize the form but not the fact itself.
but if you (synthetically) rephrase the text several times, it’ll memorize the fact
arxiv.org/abs/2309.14316
If you show a fact to an LLM in pre-training once, it’ll memorize the form but not the fact itself.
but if you (synthetically) rephrase the text several times, it’ll memorize the fact
arxiv.org/abs/2309.14316
Physics of Language Models: Part 3.1, Knowledge Storage and Extraction
Large language models (LLMs) can store a vast amount of world knowledge, often extractable via question-answering (e.g., "What is Abraham Lincoln's birthday?"). However, do they answer such questions ...
arxiv.org
November 16, 2025 at 4:59 PM
the wild part — this probably doesn’t have anything to do with AI or LLM architecture
e.g. the gospels of Matthew, Mark, Luke & John are all rephrasings of the same stories
Rabbinic teaching involved repetitive rephrasing
oral traditions in general are less about precise recitation
e.g. the gospels of Matthew, Mark, Luke & John are all rephrasings of the same stories
Rabbinic teaching involved repetitive rephrasing
oral traditions in general are less about precise recitation
Reposted by Thomas Wood
Kosmos: an AI scientist capable of 6-month (human time) work tasks
The wild 6 month claim is believable — they validated their users’ claims by finding the same work had already been done, and took that long
edisonscientific.com/articles/ann...
The wild 6 month claim is believable — they validated their users’ claims by finding the same work had already been done, and took that long
edisonscientific.com/articles/ann...
November 16, 2025 at 6:27 PM
Kosmos: an AI scientist capable of 6-month (human time) work tasks
The wild 6 month claim is believable — they validated their users’ claims by finding the same work had already been done, and took that long
edisonscientific.com/articles/ann...
The wild 6 month claim is believable — they validated their users’ claims by finding the same work had already been done, and took that long
edisonscientific.com/articles/ann...
We've come a long way
"Protect Larry Summers at all costs" said no Democrat ever.
November 16, 2025 at 7:01 PM
We've come a long way
talking to kimi k2 through my dioxus frontend and async_openai rust backend and it's pretty cool!
November 16, 2025 at 10:25 AM
talking to kimi k2 through my dioxus frontend and async_openai rust backend and it's pretty cool!
neat
ReVeal: Self-Evolving Code Agents via Reliable Self-Verification
Reinforcement learning with verifiable rewards (RLVR) has advanced the reasoning capabilities of large language models. However, existing methods rely solely on outcome rewards, without explicitly opt...
arxiv.org
November 16, 2025 at 4:42 AM
neat
Reposted by Thomas Wood
yeah, K2-Thinking is not the top coding model, but it's dealing with Rust just fine. Long horizon tasks, adjusting without doom loops, idk i think it's time to call it:
Rust is not a problem
Rust is not a problem
November 16, 2025 at 2:32 AM
yeah, K2-Thinking is not the top coding model, but it's dealing with Rust just fine. Long horizon tasks, adjusting without doom loops, idk i think it's time to call it:
Rust is not a problem
Rust is not a problem
okay going to switch to kimi k2 with opencode probably then
impressions so far, ~1 hour in
* K2 in Claude Code is slow, and doesn't have a dramatically different feel from other models
* Rust with AI seems fine, mostly. There were a couple issues, but a CLAUDE.md fixed them
* K2 in Claude Code is slow, and doesn't have a dramatically different feel from other models
* Rust with AI seems fine, mostly. There were a couple issues, but a CLAUDE.md fixed them
lettuce begin
November 16, 2025 at 3:47 AM
okay going to switch to kimi k2 with opencode probably then
Reposted by Thomas Wood
honestly, similar for me. i’ve been using K2 even though i pay for chatgpt
talk about “adaptive reasoning” in GPT-5.1, well K2 already does that well. and the personality doesn’t feel bolted on
talk about “adaptive reasoning” in GPT-5.1, well K2 already does that well. and the personality doesn’t feel bolted on
November 15, 2025 at 10:16 PM
honestly, similar for me. i’ve been using K2 even though i pay for chatgpt
talk about “adaptive reasoning” in GPT-5.1, well K2 already does that well. and the personality doesn’t feel bolted on
talk about “adaptive reasoning” in GPT-5.1, well K2 already does that well. and the personality doesn’t feel bolted on
I've got two opencode agents talking to each other in the backend right now hahahahaha
November 15, 2025 at 9:51 PM
I've got two opencode agents talking to each other in the backend right now hahahahaha
$ agentctl get agents
NAME PROJECT STATE TURNS UPTIME
supervise-robotics robotics-vla running 23 2h
build-nuclear-001 nuclear-moose running 8 45m
architect-main portfolio running 156 24h
NAME PROJECT STATE TURNS UPTIME
supervise-robotics robotics-vla running 23 2h
build-nuclear-001 nuclear-moose running 8 45m
architect-main portfolio running 156 24h
November 15, 2025 at 8:59 PM
$ agentctl get agents
NAME PROJECT STATE TURNS UPTIME
supervise-robotics robotics-vla running 23 2h
build-nuclear-001 nuclear-moose running 8 45m
architect-main portfolio running 156 24h
NAME PROJECT STATE TURNS UPTIME
supervise-robotics robotics-vla running 23 2h
build-nuclear-001 nuclear-moose running 8 45m
architect-main portfolio running 156 24h