Reading markets, AI. Ignoring the noise
Joined July 2026
- Tweets246
- Following51
- Followers67
- Likes192
Pinned Tweet
Gemini fans, look away🤓
Same prompt: "3D human anatomy atlas".
Gemini 3.8 Flash: floating dots.
GPT-6 Astra: a full human body. Muscles, bones, organs. Click any part and it tells you what it is.
Med students would love this. It runs in a browser tab.
Prompt in the first reply 👇
Gemini or Astra? Reply 1 or 2
Aimore retweeted
STOP BUILDING YOUR CLAUDE SECOND BRAIN BY HAND
Obsidian’s CEO already open-sourced the skills and four more repos finished the rest.
Most people still copy broken tutorials from last month.
The play:
Clone kepano/obsidian-skills.
Point Claude Code at your vault.
Drop any file into /raw and type ingest.
The map writes the page, links it to five existing ones and updates the rules with how you actually think.
After 90 days the search bar cites your own notes with dates.
Google cites strangers.
Your vault cites you.
The people running this don’t open ChatGPT anymore.
They open their own search.
The catch: without a source date on every ingest the vault starts inventing.
One missing date and it fills the gap with whatever sounds right.
One 5-minute setup.
Runs forever on the laptop in front of you.
Compounds for years.
Bookmark this before you open another empty chat.
Aimore retweeted
YOUR AGENT FORGETS EVERYTHING WHEN YOU SWITCH TOOLS
Living Brain keeps one persistent memory across ChatGPT, Claude, Cursor, Codex and your own agents. Write in one, retrieve in any other.
the play:
→ connect via plugin or MCP
→ save a decision or note in one tool
→ open the next tool and ask for it
→ it pulls the memory in ~464ms
> the catch
the memory lives outside the model. check what gets stored and who can access it
> what not to do
don't dump every chat into it on day one. start with decisions and project context only
one brain that travels with you beats rewriting the same context every time you switch tools
bookmark this
Aimore retweeted
Sophie Rain has shown ~$100M+ in OnlyFans gross. High-profile people with real money online are constant targets.
A recent case shows how little it takes: a user lost thousands after three days on open hotel Wi-Fi. No phishing link. No leaked seed phrase. Just a fake approval prompt that looked normal while he was on a legitimate site.
Here’s the usual play:
→ Attacker joins the same local network (hotel, Airbnb, café, or a compromised router).
→ Injects or sits in the middle of traffic.
→ You open your wallet or exchange as usual.
→ A “reconnect,” “verify,” or approval request appears.
→ You sign once. Funds move later (or immediately).
On-chain there is no undo.
How to stay uncooked:
→ Never touch a wallet or exchange on public/shared Wi-Fi. Use a hotspot or VPN.
→ Keep serious size on a hardware wallet.
→ Read every signature on the device before confirming—especially permissions and unlimited approvals.
→ Lock down your own router (change defaults, disable remote access).
Your bag is only as safe as the network you open it on.
Save this before the next trip.
Aimore retweeted
ROBOTS LEARN TO STACK BOXES IN SIMULATION BEFORE THEY EVER TOUCH A REAL ONE
Ambi Robotics put out a promo for AmbiStack and the part that got me is how they train it. The robot runs through a ridiculous number of pallets in a simulated warehouse first, and the company claims that's 10,000x faster than learning the slow way on a real floor. Then the same brain goes into an actual building and starts stacking real cartons without anyone babysitting it. The sim gets the reps, the real warehouse gets the finished worker. Which is probably why peak season stops sounding scary when you can test a thousand bad days before the real one shows up. The first half of this clip is my own cartoon version of the idea, the second half is the real thing
Aimore retweeted
🚨THE TWENTY DOLLAR CHATGPT SUBSCRIPTION WAS JUST A SMOKESCREEN
AND OPENAI REVENUE QUIETLY FLIPPED TO WALL STREET ENTERPRISE CYBER DEFENSE
Greg Brockman just confirmed OpenAI revenue officially inverted from seventy percent consumer chat subscriptions to majority enterprise contracts
The anchor deal is an unannounced partnership with JPMorgan Chase deploying Codex and autonomous cyber models to protect four trillion dollars in banking assets
While retail users argue over chatbot tone the most conservative bank on Wall Street is handing its security perimeter to frontier reinforcement learning
If you are still building consumer prompt wrappers you missed the rotation Frontier compute is no longer funded by consumer curiosity but by enterprise balance sheets locking down critical infrastructure
When the world largest bank trusts autonomous models more than human security teams what does your SaaS security moat actually protect?
Aimore retweeted
My Claude Code setup in my head vs in real life
In my head: a 14-day schedule, lead scoring, a pipeline that rates every post for "viral_fit", all of it running on $0.42
In real life: one terminal tab and a CLAUDE.md I keep forgetting to update
Who here actually has the dashboard?
Aimore retweeted
Jennifer Pan's father woke up from a coma after being shot in the face. The first thing he told police was that his own daughter was standing with the gunmen. Calm. Talking. Not tied up. She had hired them to kill both parents for $10,000.
On November 8, 2010, three men broke into the Pan family home in Markham, Ontario. Jennifer called 911 and said armed strangers tied her to a banister and shot her parents. Her mother Bich was killed. Her father Hann took a bullet to the face.
Police almost believed her. She cried in every interview, thanked the officers, played the perfect victim. She changed her story when the facts stopped matching. First she was a hostage. Then she said she hired the men to kill herself, not her parents. Detectives didn't buy it.
Small things gave her away. She was tied up yet somehow called 911 from her phone. Valuables were left untouched. A wallet full of cash, a watch, a camera in the safe. Nobody took anything.
They brought her in three times. By the third interview she broke. She admitted she hired the men through her boyfriend Daniel Wong, a drug dealer she had been dating in secret for 6 years.
Jennifer had been lying to her parents for 4 years. Told them she was in university studying to be a pharmacist. She never enrolled. When they found out and banned her boyfriend, she paid strangers to end their lives.
She was an award-winning pianist from age 4. Her parents believed she had a future. She built a fake life and hired hitmen instead.
Found guilty of first-degree murder and attempted murder. Life in prison with no parole for 25 years. Wong and two hitmen got the same sentence.
Her own father survived a bullet to the face and put her away. Would you testify against your child if they tried to kill you?
Aimore retweeted
$2.3 billion in crypto profits. Now Congress wants answers.
Reuters estimated the Trump family made at least $2.3 billion from its major crypto ventures through April 2026. Democrats take control of both chambers in January, and senior lawmakers plan to investigate the family's business dealings. Rep. Robert Garcia's message: “Lawyer up.”
A firm tied to the UAE reportedly paid $500 million for 49% of World Liberty Financial, the company behind USD1. Separately, a UAE fund used that stablecoin for a $2 billion investment in Binance. Trump later pardoned Binance founder Changpeng Zhao.
The White House denies wrongdoing. The CLARITY Act was already blocked in the Senate over concerns including conflicts of interest. For crypto, the next political fight is about who writes the rules and whose business interests come under scrutiny.
Aimore retweeted
A viral clip says Claude Opus 5.5 taught drones to hunt people. Its own telemetry says otherwise😳
Look at the chart legends: “SUPER commanded”, a person labeled “walker”, flight mode “Offboard”. SUPER is an open-source LiDAR drone project from the University of Hong Kong. Purdue’s 44-second forest demo, posted Oct 23 as built from scratch, shows the same kind of overlay, and DroneXL noticed it names SUPER too.
What the viral post gets wrong:
“Heat map” is a color-coded 3D LiDAR map, not thermal
“FPV hunter” is a research quadcopter on a person-following task
“Claude trained it”: I found no source for that
The real Claude and drones story: Andon Labs built Drone-Bench, based on its Project Pilot work with Anthropic. Models write code that makes a cheap drone find and follow a person.
They don’t train or fly anything. Opus 5.5 topped it at launch
The honest worry is dual use. Person-following is a search-and-rescue demo and a surveillance demo at once.
Video: as circulated, original source unknown
A viral clip says Claude Opus 5.5 taught drones to hunt people. Its own telemetry says otherwise😳
Look at the chart legends: “SUPER commanded”, a person labeled “walker”, flight mode “Offboard”. SUPER is an open-source LiDAR drone project from the University of Hong Kong. Purdue’s 44-second forest demo, posted Oct 23 as built from scratch, shows the same kind of overlay, and DroneXL noticed it names SUPER too.
What the viral post gets wrong:
“Heat map” is a color-coded 3D LiDAR map, not thermal
“FPV hunter” is a research quadcopter on a person-following task
“Claude trained it”: I found no source for that
The real Claude and drones story: Andon Labs built Drone-Bench, based on its Project Pilot work with Anthropic. Models write code that makes a cheap drone find and follow a person.
They don’t train or fly anything. Opus 5.5 topped it at launch
The honest worry is dual use. Person-following is a search-and-rescue demo and a surveillance demo at once.
Video: as circulated, original source unknown
Overlay breakdown: dronexl.co/2026/10/26/purdue…
TechSpot: techspot.com/news/114361-wat…
Drone-Bench: andonlabs.com/evals/drone-be…
Anthropic’s Project Pilot: anthropic.com/research/proje…
Aimore retweeted
An employee asks, "What am I responsible for?"
An owner asks, "What happens if nobody takes responsibility?"
The second question changes how you see a business.
Aimore retweeted
THE MACH TWO DOGFIGHT IS NO LONGER FOUGHT BY HUMANS
AND HUMAN BIOLOGY JUST BECAME A FATAL DESIGN FLAW IN SUPERSONIC COMBAT
Saab just revealed that proprietary AI agents completed live flight trials autonomously controlling the Gripen E multirole fighter jet
During live beyond-visual-range missile scenarios the agent flew the supersonic jet with zero human hands on the flight controls
Trained on millions of tactical combat hours inside Saab virtual flight academy the model was flashed directly into frontline software-defined avionics
Human pilots blackout at nine Gs and process telemetry in hundreds of milliseconds while reinforcement learning calculates missile intercept vectors at clock speed
If autonomous agents can command supersonic kinetic warfare at Mach 2 the debate over trusting AI in mission-critical software is officially over
Aimore retweeted
I asked Claude to build me a beach resort and I didn’t expect it to look this real 🏖️🤯
A whole coastline with overwater bungalows, a pool, a long wooden pier and palm trees swaying over turquoise water. The camera just keeps flying and I keep watching.
Would you actually stay here?
Aimore retweeted
A VISITOR DROPPED A WHOLE STACK OF PAPERS. WHAT THIS ROBOT DID NEXT FROZE THE ROOM!
The robot was handing out flyers at a busy expo when a visitor dropped the entire pile on the floor. It silently bent down, collected every sheet and handed the stack back as if it had done this a thousand times.
For a robot that’s a huge task. It has to spot thin paper on a smooth floor, judge the distance, keep its balance while bending and grip the sheets without crushing them.
Today it’s flyers. Tomorrow it’s handing you your keys and clearing your table.
The future of robots won’t start with a big dramatic moment. It will start with a machine quietly fixing your small mistake.
Aimore retweeted
7.9% to 48.1% on MATH-500. No training run
MIT CSAIL researchers gave GPT-6 Astra the raw weights of Qwen3.5-0.8B and asked it to make the model better at math, without gradients.
Astra found the cause of the repetition loops that Qwen’s own model card warns about and edited 0.016% of the weights. MATH-500 went from 7.9% to 48.1%. IFEval went from 21.0% to 62.3%, about 3x.
It could also tell what a small game network was trained to do just from its weights, and it wrote small game networks from scratch that beat PPO baselines trained for 100M steps in all three games.
My take: this is not “AI learned math”. The gain came mostly from stopping repetition loops. The real result is an LLM finding a bug in another model by reading its weights.
Caveats: authors’ blog post, not peer reviewed. One model, no independent replication of the edit yet. On Gemma 3 270M there was no established math gain. Critics say a human ML researcher would have made the same fix.
Which matters more: LLMs reading weights or editing them?
7.9% to 48.1% on MATH-500. No training run
MIT CSAIL researchers gave GPT-6 Astra the raw weights of Qwen3.5-0.8B and asked it to make the model better at math, without gradients.
Astra found the cause of the repetition loops that Qwen’s own model card warns about and edited 0.016% of the weights. MATH-500 went from 7.9% to 48.1%. IFEval went from 21.0% to 62.3%, about 3x.
It could also tell what a small game network was trained to do just from its weights, and it wrote small game networks from scratch that beat PPO baselines trained for 100M steps in all three games.
My take: this is not “AI learned math”. The gain came mostly from stopping repetition loops. The real result is an LLM finding a bug in another model by reading its weights.
Caveats: authors’ blog post, not peer reviewed. One model, no independent replication of the edit yet. On Gemma 3 270M there was no established math gain. Critics say a human ML researcher would have made the same fix.
Which matters more: LLMs reading weights or editing them?
Source: MIT CSAIL authors (Takehi, Teoh, Zhang, Khattab), “Sparks of Weight-Space Programming”, blog post, Nov 2, 2026. rikiyat.github.io/blog/weigh…
Note: Astra beat the 100M-step PPO baselines in all three games, but trailed 500M-step PPO in Seaquest. Coverage: VentureBeat, Nov 2.
Aimore retweeted
A fruit fly brain reviewed code. Its author’s own README says a spreadsheet could do it better🧐
Nick Nisi (WorkOS) wired a simulated male fruit fly connectome, 176,422 neurons, into code review. A helper AI answers 30 yes/no questions about each change, and the answers become a smell. If the smell makes the fly stick its tongue out, it approves.
Main test, 301 changes it never saw:
Fly: 88.3% balanced accuracy (coin flip: 50%)
Plain logistic regression on the same smell: 94.8%
Harder test, 24 tickets that change behavior on purpose:
The fly approved 0 of 24 correct changes and caught 22 of 24 mistakes. It rejects whatever changes behavior
The classifier did equally badly: both scored 45.8%
The author also switched off 2 of the 30 questions, because one of them alone sorted 296 of 301 changes. That would have made the fly a middle man.
My take: the fly isn’t the story. The honesty is, and the lesson that “behavior changed” is not “wrong”: a reviewer needs the intent.
Credit: Nick Nisi
