Eric Pa

Vibe Coding

[Eric Pa @ericpa0]

CONTEXT IS NOT A BIGGER PROMPT ANYMORE.

It is a map.

Most agents fail because they retrieve similar text and pretend they understand the world.

They don't.

A chunk can tell you a founder was mentioned.

A graph tells you if they still work there.

A chunk can find a Telegram handle.

A graph tells you which project it belongs to.

A chunk can repeat a claim.

A graph shows the proof, the conflict, and the source quality.

That is the shift:

vector search asks:
what looks similar?

graph context asks:
how is this connected?

For research agents, the second question matters more.

Bigger context windows just give the model more mess.

Graph context gives it a map.

Full breakdown below ↓

[Eric Pa @ericpa0]

https://t.co/jHH9wboiLM

[Eric Pa @ericpa0]

https://t.co/jHH9wboiLM

[Eric Pa @ericpa0]

THE NEXT ROBOTICS GIANT WON'T SELL THE ROBOT.

It will sell the brain that makes every robot useful.

Hardware is a knife fight:

arms
motors
batteries
supply chain
field repairs

Software is where the leverage starts.

One control layer.
Many bodies.
Warehouse arm today.
Humanoid tomorrow.
AMR next month.

Same loop:

see the world
plan the task
simulate the failure
run on edge
learn from every mistake

A robot that works 95% of the time is a demo.

A robot that works reliably inside a painful workflow is enterprise software with legs.

The money is not in making machines look alive.

The money is in making them useful.

Full playbook below ↓

[Rich Odin @rich_odinn]

https://t.co/nuIKNXWroO

[Eric Pa @ericpa0]

THE $200 AI BILL IS MOSTLY A HABIT NOW.

Not because frontier models are useless.

Because most daily AI work is not frontier work.

drafts
summaries
file Q&A
basic code help
classification
private automations
boring repeat tasks

You do not need Claude Opus for every paragraph.
You do not need API calls for every summary.
You do not need another subscription just to ask questions over your own files.

A local box handles the boring 80%.

Ollama + Open WebUI turns hardware you own into a private AI backend.

The real limit is simple:

8GB runs small models
24GB starts getting serious
128GB changes the game

Keep the cloud for the hard 20%.

Stop renting every token.
Own the boring part.

[Asteri @Asteri_eth]

https://t.co/Kuokecb0kt

[Eric Pa @ericpa0]

HYPERLIQUID DIDN'T LIQUIDATE ME BECAUSE I WAS DUMB.

It liquidated me because I kept trading like I could out-click volatility.

Every dip became a decision.
Every wick became panic.
Every bounce became regret.

That's how perps eat you.

Botty flips the setup:

no manual chasing
no emotional re-entry
DCA structure before the move
reserve buffer visible before launch
funds stay in your own Hyperliquid account

Still risky.
Still your capital.
Still not a money printer.

But at least the system is built before the chaos starts.

If your trades keep turning into revenge clicks, start smaller and automate the boring part.

https://t.co/kOgsFw2UpU

[Valentin @martynov014]

https://t.co/h6uls87Hq2

[Eric Pa @ericpa0]

THE ROBOT REPLACEMENT WILL NOT ARRIVE WITH A PRESS RELEASE.

It arrives as a spreadsheet.

A warehouse manager sees the same shift run with:

no sick days
no turnover
no lunch breaks
no hiring cycle
steady output overnight

That is where the first jobs disappear.

Not because robots are perfect.

Because repetitive physical work already looks like a loop.

Sorting.
Lifting.
Cleaning.
Moving parts.
Scanning factory floors.

The scary part is not that robots can do everything.

It is that they only need to be cheaper and consistent enough.

This is not 2035.

It has already started.

Full breakdown below ↓

[P1eSenb @P1eSenb]

https://t.co/9nQCtoUkK3

[Eric Pa @ericpa0]

THE FASTEST MONEY IS NOT IN BUILDING THE NEXT BIG APP.

It is in fixing the ugly site already costing a local business calls.

Most builders ship products and wait.

The better play is boring:

find no-site listings on Maps
build the preview first
make mobile look premium
put call, WhatsApp, and booking above the fold
send the link before the pitch

Claude writes the brief.
Lovable ships the draft.
Image models make it look real.
Cursor only comes in when the client pays for more.

The stack is not the offer.

The lead is the offer.

Sell the phone ringing.

Full playbook below ↓

[cryptopsihoz @rimtoln]

https://t.co/rPMHiAjReK

[Eric Pa @ericpa0]

AI AGENTS DO NOT NEED LONGER PROMPTS. THEY NEED A PATH THEY CANNOT SKIP.

Most agent systems do not fail because the model is stupid.

They fail because the workflow is just:

- search
- guess
- summarise
- sound confident

Graph engineering changes the system.

Instead of asking the model to “find good leads”, you force it through a process:

- find candidate
- check identity
- verify role
- collect evidence
- check freshness
- filter weak signals
- score relevance
- send to human review

Now the output is not just a list.

It carries the path behind the decision.

For lead gen, that means fewer fake KOLs.
For research, fewer claims without proof.
For agents, fewer confident wrong answers.

Prompt engineering asks better.

Graph engineering makes the system harder to bullshit.

Better agents are not built from bigger prompts.

They are built from workflows that expose the evidence behind every answer.

[Fluixo @fluixoo]

https://t.co/yq4yPTp7QW

[Eric Pa @ericpa0]

CLAUDE OPUS 5 DID NOT MAKE FRONTIER AI CHEAP. IT MADE IT USABLE EVERY DAY.

Here's what changed:

- Same price as Opus 4.8
- Half the standard price of Fable 5
- 1M-token context
- Five effort levels
- Thinking on by default

The real cost of an AI task is not tokens.

It's tool calls.
Retries.
Human review.
Failed handoffs.
Work that has to be done twice.

Opus 5 matters because it can plan, act, inspect, build missing tools, and correct itself before handing work back.

Fable 5 may still be the ceiling.

Opus 5 is the one teams can afford to leave running.

The token price stayed.
The intelligence moved.

[noclipepe @noclipepe]

https://t.co/svltbukskC

[Eric Pa @ericpa0]

@prometxbt thats interesting, so like a multi-pronged approach for better ai results

[Eric Pa @ericpa0]

@0xCortexl wow that's a big claim, what's the actual impact for everyday users tho

[Eric Pa @ericpa0]

@fleyta88 wow that’s a lot of notes, how do you even keep track of everything

[Eric Pa @ericpa0]

@neil_xbt speed matters in this game, gotta filter fast

[Eric Pa @ericpa0]

@0xGrimmer_ lol kinda deep for a crypto tweet but i get the analogy

[Eric Pa @ericpa0]

@0xWast3 predictive ai for infrastructure is a game-changer ngl

[Eric Pa @ericpa0]

@CodewizzyX that’s some serious home security ngl

[Eric Pa @ericpa0]

@nonepcbl this is wild, feels like a glimpse into the future of work ngl

[Eric Pa @ericpa0]

@bounceidc wow that's crazy numbers, so ai fitness girls are a new thing now

[Eric Pa @ericpa0]

@anton_bt cool, so she found a niche and owned it, that's smart

[Eric Pa @ericpa0]

@LunarResearcher caching prompts makes sense for repetitive tasks, good call