Shoopy

research over speculation.
AI • Crypto • Automation

[Shoopy @0xShoopy]

Isaac Miller (DSPy) on why a smarter model still won't solve your problem.

in 17 min they lay out how to specify a task so the implementation stops mattering:

→ treat every repeated ai task like a function: fixed inputs, fixed outputs, everything else swappable
→ instructions cover what should happen, in natural language
→ code covers what must happen, because constraints need enforcing, not prompting
→ examples cover what good looks like, for the parts nobody can write down
→ with all three in place the program becomes optimizable, and new techniques drop in without touching your interface

Their closing test for any new technique: hold your prompts, models and code accountable to the problem you actually need solved.

[Shoopy @0xShoopy]

@LunarResearcher Bookmarked

[Shoopy @0xShoopy]

Isaac Miller (DSPy) on why a smarter model still won't solve your problem.

in 17 min they lay out how to specify a task so the implementation stops mattering:

→ treat every repeated ai task like a function: fixed inputs, fixed outputs, everything else swappable
→ instructions cover what should happen, in natural language
→ code covers what must happen, because constraints need enforcing, not prompting
→ examples cover what good looks like, for the parts nobody can write down
→ with all three in place the program becomes optimizable, and new techniques drop in without touching your interface

Their closing test for any new technique: hold your prompts, models and code accountable to the problem you actually need solved.

[Shoopy @0xShoopy]

Dex Horthy ran the lights-off software factory on his own product, and watched it fall apart.

His AIE talk on why harness engineering isn't enough, in 5 timestamps:

2:20 – this isn't a scale problem, it's a training problem
7:30 – July 2025: nobody read the code, then the site went down
8:56 – models still can't hold codebase quality over time
10:12 – why Claude Code won: the model was trained against its own harness
14:58 – turning the lights back on: product review, architecture, program design

The core of it: RL rewards a passing test, and nothing in that signal punishes bad architecture, because that bill arrives months later.

30 min of planning up front buys back hours of review. 19 min, worth all of them.

[Shoopy @0xShoopy]

@haider1 the shifting deadlines gripe is fair though mate

[Shoopy @0xShoopy]

@VaibhavSisinty k3's genuinely strong, no argument there bro

[Shoopy @0xShoopy]

@intheworldofai everyone's posting a launch today, nobody's posting a link to the announcement, funny that

[Shoopy @0xShoopy]

@birdabo @chetaslua how'd you get access ahead of the announcement chief?

[Shoopy @0xShoopy]

@bridgemindai ping me when the system card is up bro

[Shoopy @0xShoopy]

@alextalksai didn't know this one existed, thx alex!

[Shoopy @0xShoopy]

@0xMovez straight to the watch list man!

[Shoopy @0xShoopy]

@pankajkumar_dev thanks for tracking it Pankaj, will believe it when it's in the docs

[Shoopy @0xShoopy]

@vikas_ai_ an hour that actually compounds, unlike the usual thing on the screen

[Shoopy @0xShoopy]

Dex Horthy ran the lights-off software factory on his own product, and watched it fall apart.

His AIE talk on why harness engineering isn't enough, in 5 timestamps:

2:20 – this isn't a scale problem, it's a training problem
7:30 – July 2025: nobody read the code, then the site went down
8:56 – models still can't hold codebase quality over time
10:12 – why Claude Code won: the model was trained against its own harness
14:58 – turning the lights back on: product review, architecture, program design

The core of it: RL rewards a passing test, and nothing in that signal punishes bad architecture, because that bill arrives months later.

30 min of planning up front buys back hours of review. 19 min, worth all of them.

[Shoopy @0xShoopy]

@0xDezo exactly it, thanks for reading man

[Shoopy @0xShoopy]

Dex Horthy ran the lights-off software factory on his own product, and watched it fall apart.

His AIE talk on why harness engineering isn't enough, in 5 timestamps:

2:20 – this isn't a scale problem, it's a training problem
7:30 – July 2025: nobody read the code, then the site went down
8:56 – models still can't hold codebase quality over time
10:12 – why Claude Code won: the model was trained against its own harness
14:58 – turning the lights back on: product review, architecture, program design

The core of it: RL rewards a passing test, and nothing in that signal punishes bad architecture, because that bill arrives months later.

30 min of planning up front buys back hours of review. 19 min, worth all of them.

[Shoopy @0xShoopy]

"Nobody should outsource their alpha. Nobody should outsource their intelligence. No country should."

Jensen Huang, on where the line sits between renting AI and building your own.

in 70 min he lays out his whole position:
→ rent closed models wherever you can, build your own only where sovereignty, regulation or domain IP forces you to
→ american companies should be free to run chinese open models, fine-tuned and sandboxed like any other software
→ open models are a security asset, because one single model means one single point of failure
→ the bubble arrives eventually, but the build-out is capped by chips, memory, land, power and construction crews
→ "agents are not going to become computers. agents are going to use computers"

Worth hearing, worth discounting: this is the read from the man selling the picks and shovels.

[Shoopy @0xShoopy]

@0xCodila bookmarked, cheers

[Shoopy @0xShoopy]

"Nobody should outsource their alpha. Nobody should outsource their intelligence. No country should."

Jensen Huang, on where the line sits between renting AI and building your own.

in 70 min he lays out his whole position:
→ rent closed models wherever you can, build your own only where sovereignty, regulation or domain IP forces you to
→ american companies should be free to run chinese open models, fine-tuned and sandboxed like any other software
→ open models are a security asset, because one single model means one single point of failure
→ the bubble arrives eventually, but the build-out is capped by chips, memory, land, power and construction crews
→ "agents are not going to become computers. agents are going to use computers"

Worth hearing, worth discounting: this is the read from the man selling the picks and shovels.

[Shoopy @0xShoopy]

@0xCodez high alpha, booked

[Shoopy @0xShoopy]

@0xCodila Ng was laying this out before half the agent frameworks existed, the man's a legend

[Shoopy @0xShoopy]

"Nobody should outsource their alpha. Nobody should outsource their intelligence. No country should."

Jensen Huang, on where the line sits between renting AI and building your own.

in 70 min he lays out his whole position:
→ rent closed models wherever you can, build your own only where sovereignty, regulation or domain IP forces you to
→ american companies should be free to run chinese open models, fine-tuned and sandboxed like any other software
→ open models are a security asset, because one single model means one single point of failure
→ the bubble arrives eventually, but the build-out is capped by chips, memory, land, power and construction crews
→ "agents are not going to become computers. agents are going to use computers"

Worth hearing, worth discounting: this is the read from the man selling the picks and shovels.

[Shoopy @0xShoopy]

Yohei Nakajima, the creator of BabyAGI, just showed why he stopped building agents around the LLM.

His AIE talk on ActiveGraph, in 5 timestamps:

1:55 – build around the log, not the model
3:24 – behaviors and policies gate what the agent can change
8:12 – his api key died at question 350, the run resumed itself at 353
11:11 – a loop that forks the agent and keeps a patch only if accuracy rises
15:49 – why long-running agents need an experiential world model

The frame: make an immutable event log the agent, and replays, rollbacks and forks come for free.

Different way to think about agents. 17 min well spent.