You can just build things
Virginia
Joined January 2014
- Tweets4.4K
- Following1.1K
- Followers247
- Likes55.9K
Pinned Tweet
We live in uncertain yet incredible times.
Since the "ChatGPT moment" two years ago, I've seen the writing on the wall as many of you have that AI will change humanity in a powerfully non-negotiable way.
Our path forward isn't to resist this change—it's to harness these extraordinary advances to elevate human potential.
Learning is the cornerstone business practice of all human development and advancement. To progress ourselves, we need to enhance and overhaul our learning system. The signal for this change has never been louder than it is today.
In my work "Learn to Adapt" I outline how we can adapt the American education system to advance ourselves in this incoming decade of AI transformation.
Read the full blueprint here -> tinyurl.com/LearntoAdapt
I'm not just sharing this as food for thought—I'm directly inviting you to join a mission through the @AdaptEducation_ project.
This rapid-development research project is funded through blockchain technology (Solana - $ADAPT). As we progress, every bit of funding drives research into methods, tools, curriculum development and more that will directly shape how AI transforms education.
Promising experiments in AI-driven education are already happening today, and we need more of them to ensure scaled and operational success across the entire US. We must fail fast and push the edge of innovation to adapt our system quickly enough to avoid the potentially catastrophic side effects of this AI revolution.
I'll continue to share my thoughts here on X about why and how we can adapt our learning system, exactly what this project will achieve, and more.
Micah retweeted
is Opus 5.5 nerfed?
I just had Gemini 4 Argon analyze 1,400 Claude Code sessions and it found that on November 3, something definitely happened.
the model name didn't change, but the way it does reasoning took a sharp turn.
Micah retweeted
Report to your nearest freedom reeducation camp
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
Codex sucks. Hermes and Claude Code are the only two serious harnesses for desktop agents IMO currently.
(Have not explored deepseek yet)
We absolutely have to provide an answer for the total lack of meaningful employment in the US. Yes.
Silicon Valley keeps celebrating how many jobs AI can eliminate.
Young Americans aren't nearly as excited.
64% of Americans ages 18–29 consider AI replacing jobs a serious threat to their financial security.
And the concern crosses political lines.
76% of young Democrats.
61% of young Republicans.
61% of young independents.
The people building AI are talking about productivity.
The people entering the workforce are worried about having a paycheck.
Micah retweeted
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
Micah retweeted
Estoy bastante de acuerdo con lo que ha dicho Andrej Karpathy recientemente. Con la salida inminente de Qwen 4, Fable 5.5 y Astra 6.2, parece que cada vez celebramos más los modelos con un mayor número de parámetros.
Es cierto que tenemos que celebrar que los modelos sean cada vez más inteligentes, pero creo que también deberíamos centrarnos en conseguir modelos con menos parámetros que sean capaces de igualar a otros mucho más grandes.
Para mí, esa es la verdadera evolución de la IA.
Micah retweeted
we have a winner for the most absurd AI startup ever
this Peter Thiel backed company uses AI robots to stab fish in the brain so they taste better and last up to 3x longer
it's called Shinkei and it modernizes a centuries-old Japanese technique
dozens of Michelin-star restaurants are already lining up to buy the fish
here's how it works:
> the problem is when a fish gets caught, it usually just suffocates on the boat
> that stress floods the meat with hormones and lactic acid, which hurts the flavor and texture
> the old Japanese fix for this was spiking the fish through the brain right after the catch, which stops all that
> but doing it by hand is impractical for most American fishing crews
> so Shinkei built a fridge-sized robot that sits right on the boat's deck
> the crew feeds in a fish and AI identifies the exact species, so it knows exactly where the brain and gills are
> the robot then drives the spike through the brain and cuts near the gills in about 7 seconds
> the fish goes into an ice slurry so the blood drains out
> so the fish doesn't sit there stressing out, and that tasty flavor and texture is preserved
example number 5,392,417 of how AI can be applied to literally anything
ALT A gloved fisherman on a boat feeds a live salmon head-first into the steel intake of a white Shinkei Systems robot, with the ocean behind.
Micah retweeted
Mind you its a data center that
- Is covering 100% of its own power consumption
- Uses zero water for cooling, just bathroom and kitchen
- Offers a 120% water stewardship pledge (Net gain of water in the county)
- Provided housing to construction workers to prevent a local temporary rent hike
- Investing in the local school system
- Was going to pay for the entire County's budget for 7 years
And it still wasn't enough. There just is zero way to overcome the negative PR/NIMBY
Micah retweeted
A quarter of a century ago, USMC GEN Paul Van Ripper did this exact strategy.
In 2002, he was tapped by the Pentagon to play a adversarial force (red team) modeled after Iran during the $250 million Millennium Challenge 2002 wargame.
General Van Ripper refused to hold back. Cleverly using low-tech, asymmetric tactics, he launched a surprise strike that sank 16 to 19 U.S. warships including an aircraft carrie) within minutes.
Van Rippers tactics were simple: motorcycle couriers, light signals, and coded signals from mosque minarets.
Unhappy with the results, organizers halted the exercise, refloated the ships, and altered the rules to script a Blue Team victory.
General Van Ripper resigned from his advisory role, but spoke out in protest:
"Nothing was learned from this. A culture not willing to think hard and test itself does not augur well for the future."
The U.S. military produced an exercise where "the result was preordained", and not a real test of America's ability to invade a hostile nation.
He was prophetic a 24 years later.
BREAKING: Iran is now completely shifting toward low tech guerrilla warfare and asymmetric tactics in the Strait of Hormuz, deploying speedboats armed with RPGs and magnetic limpet mines to choke off crude shipments, per The Wall Street Journal.
The cost imbalance is staggering: an RPG 7 round costs roughly $400, and limpet mines are described as “cheap and easily available.” Yet a single attack can disable a tanker, as seen when the MT Louise 1 was struck by two devices off Oman. The target is a multimillion dollar asset, with new VLCCs costing around $130 million and resale prices reaching $235 million.
More than 35 incidents were recorded in October, the highest monthly total of the war, with at least five vessels hit in early November, per UKMTO.
Micah retweeted
New challenge unlocked: keeping the laser fed
Only drip feeding it orders to start but man it is FAST. Chewed through a full 96x240 plate in less than 20 minutes
Micah retweeted
human-in-the-loop
└ human-on-the-loop
└ human-near-the-loop
└ human-cc'd-on-the-loop
└ human-legally-liable-for-the-loop
Sounds interesting enough to test out. @bcherny what do you think about adding to claude?
Compaction is a mistake.
When an agent's context fills up, tools like Codex and Claude Code compact it:
1. The model prefills the entire session again
2. It writes a summary of everything so far
3. A new session starts from that summary
I measured this on a real ~400k-token session with Opus 5.5:
→ 38 s and $1.99 when the prompt cache has expired
→ 29 s and $0.17 when it hasn't
And the summary is lossy. The exact error message, the file path, the one number you needed is often gone.
So I do something new for Mainplane
Context retrieval works the other way around. Mainplane keeps the full record of every session as a file. When the context fills up, a new session starts right away with one instruction: "You are continuing session X. Retrieve key context from it as you go."
The agent searches the old session and copies the exact lines it needs, when it needs them.
Same session, same model: 15 s and $0.10 to get back to work. It does not depend on a cache hit, and nothing is paraphrased.
This is the bitter lesson applied to agent memory. Don't hand-design what the model should remember. Give it the full record and good tools, and let it decide.
Mainplane is open source (MIT). Link in the first comment.
Micah retweeted
“Commander Cody, have you completed your fiscal year Galactic Republic Cyber Awareness Training?”
“I have not general.”
“You should probably get on that.”
“Yes General.”
