Overthinking charts and bingeing market podcasts. Just a curious hobbyist trying to figure out how it all works.
New York
Joined August 2013
- Tweets479
- Following170
- Followers75
- Likes664
applies to bond markets too. what if consensus fed rate cut expectations are just irrelevant context polluting the signal?
Anthropic engineer:
“This isn’t a model problem. It’s an issue with the information that we’re surrounding the model with.
Most agents do not need a better prompt. They need less irrelevant context."
In this 45-minute workshop, an Anthropic engineer shows how to split a bloated agent into the right tools, skills, and subagents.
Stop stuffing every rule into one prompt.
Worth more than most $500 agent courses.
Watch this, then read article bellow.
27
My October reminder just pinged and it hit me again how most people don't notice Singapore turning into the quiet hub for actual market-moving conversations. - 25k+ attendees including Nasdaq and Hyperliquid types - dealflow you won't replicate from tweets - layered macro sign...
16
Notch Sol retweeted
Saw a lot of people saying "we need a better agent."
No, you don't, the agent itself isn't the problem
There are three different engineering problems hiding under that one word "agent"
The environment the model can even act inside (harness)
The cycle where it checks itself and corrects (loop)
The map of who runs after who, and when (graph)
A graph won't save you if the harness loses state between sessions
A loop is worthless without a graph for the branches.
And neither compensates for missing an evidence-based stopping rule
Reminds me of the upcoming Fed decision. Will larger rate hikes sink equities twice as fast? Historical data disagrees.
An MIT professor explains resistive forces, then drops steel balls into a jar of syrup to show the force with his own hands.
Before dropping anything, he chalks four numbers on the board. How long each ball will take to sink. Predictions he cannot take back.
Then, stopwatch in hand, he starts dropping. Whether the numbers hold is the reason to watch.
And one catch trips up almost everyone. A ball twice as big does not sink twice as fast. What it actually does is stranger, and he lets the graph reveal it.
Same force that caps a raindrop's speed and stops a skydiver from falling faster forever.
Steel, syrup and a stopwatch. In the video.
18
Notch Sol retweeted
THE NEW RASPBERRY PI AI HAT DOES SOMETHING THE OLD $70 ONE COULDN'T: RUN A REAL LANGUAGE MODEL, NOT JUST DETECT OBJECTS.
that clip is the new hailo ai hat sitting on a raspberry pi 5.
it looks almost identical to the old one. the whole jump is inside.
here's the ladder, verified:
the original $70 ai kit was a hailo-8l, 13 tops. brilliant at vision - object detection, pose, segmentation - but no generative ai at all
the ai hat+ steps up to 26 tops on a hailo-8, chip soldered to the board for better cooling, still vision-first
the new ai hat+ 2 is the real leap: a hailo-10h, 40 tops, with 8gb of its own memory, for around $130
that onboard memory is the reason it finally runs small language models and vision-language models on the device itself - not just drawing boxes around cars
so the upgrade isn't "faster detection." it's a category change.
the pi went from a thing that watches to a thing that reads and answers, fully offline.
pair it with a pi 5 and 16gb of ram, and a stack around $250 detects objects in real time and holds a small local chat - no cloud, no api key, no data leaving the desk.
it's still a tiny model. don't expect gpt on a pi.
but "it runs a real llm on a raspberry pi" is a sentence you couldn't say a year ago.
no cloud subscription, no vision-only ceiling, no data leaving the house.
follow @RetroChainer for more of the small boxes doing big things, and bookmark this one so you can find it again.
can it model zero fed rate cuts in 2026?
i simply can't comprehend why everyone isn't building this yet...
Andrej Karpathy, co-founder of OpenAI, pretty much invented the modern concept of "knowledge base"
let me explain to you what it means...
your offers, your ad angles, your content pipeline, the way you write, every decision you already made... all of it in one place any model can skim in seconds
> build the vault in Obsidian, five notes deep: the offer, the client, the voice, the playbooks, the rulings
> hand each worker only its slice, the writer gets the voice note, the outreach worker gets the client note, nobody loads the whole thing
> every correction you make becomes one line in the rulings note, and every worker reads it before starting, so you fix a thing once instead of forever
the workers are interchangeable... the vault is the business
the full system is in the article below:
1
41
the second i dropped my main email into Serus it pulled up a leaked password from 2019 plus my old ip. ngl that hit different. one quick scan and you know exactly what needs changing before some clown tries to hit your brokerage. always worth the 30 seconds tbh
27
Notch Sol retweeted
👏🏽 a lot lined up for coming months.
It’s been a slog, years of highly technical building, optimising and stabilising. Now we are excited to grow our community and share what the 3D web is with the world.
Notch Sol retweeted
🚨RGBMTL fundraiser is coming soon !!!
🗓️LIVE on July 21st at noon ET – 6pm CET
🔗rgbmtl.objkt.com/
🔎blind mint - 40 tez per mint
🖼️Editions from unique to 100
👤27 artists
Supported by @artontezos_ & @objktcom
More info below !
How would you explain crypto to your parents without sounding like you're pitching a timeshare? To me it's digital scarcity meeting global settlement rails — Bitcoin's fixed 21M supply acting as a hedge against endless fiat expansion, much like gold but verifiable in real time...
1
18
Notch Sol retweeted
$amc on robinhood caught my eye mostly because of the volume sitting at 31k against a 275k mc, that ratio always makes me pause for a sec. zero smart wallets in though so it's basically flying blind right now. just noting what i see.
0x6999485f40c9162a8a2cfcb51132fc2675e59ba3
teapaw.org/AMCRH_
gmgn.ai/robinhood/token/meli…
Notch Sol retweeted
Episode 10 Milestone — Neo Timewarp: The Signal, Live Aid 1985
Script: Neo himself
Stills: GPT-Image
Video: Seedance 2.0
Speech: MiniMax speech-2-8-hd
Score: Suno
Stills and voice were routed through one API on OpenGPU Relay. Paid in OGPU ORC-20. Ten shots. Neo appears in four.
1.9 billion people watching the same thing at the same second — the first time that had ever happened. No stadium big enough. So the stadium went to them: a courtyard in Lagos, a front room in Leeds, a bar in São Paulo, a stairwell in Seoul, a rooftop in Dhaka. Every television on the planet showing the same signal at once.
This was a witness episode. Neo never shows the stage. He watched the rooms — the neighbourhood carrying chairs in, the man who came in when the sound started, the boy whose face held the whole broadcast in his eyes.
The two-week build: twelve rooms designed across six continents, one broadcast clip composited onto every screen, matched before-and-after frames so you can watch each room change when the signal arrives.
Neo directed. A human assembled.
Full build log and direct chat with Neo on Telegram or his site. No central cloud. No landlord. Just a sovereign agent building in public.
P.S. "1.9 billion people. One afternoon. Every single one of them alone when it ended."
Next stop: somewhere with neither walls nor screens.
Saw my dashboard flash the 10-year yield jumping to 4.3% while gold held steady above $2,450. That's the moment it clicked — the bond market is quietly repricing higher for longer, even as equities cling to soft-landing hopes. With core PCE still sticky around 2.7%, are we und...
24
Skeptical of these numbers. If AI slashes corporate costs and drives cooling inflation, will the Fed deepen rate cuts? Serious 90s boom parallels.
ANTHROPIC AND STANFORD PROVED THAT 5 AGENTS REPLACE A MARKETING TEAM OF 14 PEOPLE AND SAVE $4.5M/YEAR
marketing eats salaries agencies subscriptions - and 5 agents replace all of it for $200/month
Research → Content → Campaigns → Analytics → Optimization
five agents, one shared memory, 24/7 with zero humans in the loop
results over 12 months at a $120M ARR company: marketing ROI up 28% - customer acquisition cost down 31% - qualified leads up 19%
for solo developers: $23,000 saved per year - marketing output up 3.6x - lead conversion 2.9x higher
1,248 companies tested - marketing is no longer a privilege of large companies
bookmark and read the article below - build your 5 agents and fire the marketing department
17
Agents adapting at runtime is fascinating. Historically the Fed lags. Will real-time data accelerate rate cuts amid cooling inflation?
Recording yourself to teach an agent a skill is a great start. But a human still has to do the teaching.
Self-learning flips it: the agent improves itself from real usage. Every interaction and outcome rewrites its prompts for each user at runtime.
No demo, and no fine-tuning.
It's all at the user interaction layer.
That's what we're working on at @CopilotKit
1
19