it's f*cking insane
alibaba just open-sourced the AI code reviewer that caught millions of bugs for their own devs
this repo reviews your code like a senior dev: git diff in, line-level comments out
open-code-review, a CLI that plugs into Claude or Codex:
> tens of thousands of alibaba devs, millions of bugs caught before it went public
> code picks the files, bundles them and matches the rules, the LLM only does the thinking
> every bundle gets its own sub-agent, so big PRs don't get half-reviewed
> built-in rules for null pointers, thread-safety, XSS and SQL injection
> vs Claude Code on the same model: higher precision, ~1/9 of the tokens (their own benchmark)
> delegate mode runs it on your agent's own model, no extra API key
50,000+ stars on github
your PRs are about to get roasted xd
Show more
this is f*cking dangerous
point it at any app, site or game and your AI agent builds your own version of it
Craft Anything, open source, pure Rust:
> study -> build -> check: it learns how the target works, builds it, then shows what doesn't match
> clean room by default: it never reads the original's code, only what you can see from outside. deep mode looks inside
> runs on your own Claude Code login. Codex, Copilot and others through ACP
> 320 tools, each one a CLI command and an MCP tool, so any agent can drive it
still early alpha, and it already feels unfair
my PokeTide bros tested it on my own game first xd
Show more
it's f*cking insane
alibaba just open-sourced the AI code reviewer that caught millions of bugs for their own devs
this repo reviews your code like a senior dev: git diff in, line-level comments out
open-code-review, a CLI that plugs into Claude or Codex:
> tens of thousands of alibaba devs, millions of bugs caught before it went public
> code picks the files, bundles them and matches the rules, the LLM only does the thinking
> every bundle gets its own sub-agent, so big PRs don't get half-reviewed
> built-in rules for null pointers, thread-safety, XSS and SQL injection
> vs Claude Code on the same model: higher precision, ~1/9 of the tokens (their own benchmark)
> delegate mode runs it on your agent's own model, no extra API key
50,000+ stars on github
your PRs are about to get roasted xd
Show more
this is f*cking dangerous
point it at any app, site or game and your AI agent builds your own version of it
Craft Anything, open source, pure Rust:
> study -> build -> check: it learns how the target works, builds it, then shows what doesn't match
> clean room by default: it never reads the original's code, only what you can see from outside. deep mode looks inside
> runs on your own Claude Code login. Codex, Copilot and others through ACP
> 320 tools, each one a CLI command and an MCP tool, so any agent can drive it
still early alpha, and it already feels unfair
my PokeTide bros tested it on my own game first xd
Show more
the best pokemon game in the world just got f*cking co-op
PokeTide v1.42 is here, and it's 100% free
MEGA
> 21 new Mega Evolutions: Rayquaza, Primal Groudon, Latias, Latios, Lucario, Diancie, Absol, Gardevoir, Gallade and more
> they evolve right in the middle of a fight, the sprite changes and you keep hitting
> new Mega Mode: your run starts already Mega
> Mega Forms collection, sorted by region
CO-OP
> closed beta is live: you and a friend on one map
> clear the waves together, then take the boss down together
> testing it with the first players right now, more of you soon
DUEL
> matchmaking 1v1: you and a random player on the same seed. who's stronger?
> win duels, earn MMR, climb
runs free in your browser and on your phone
which Mega are you taking into co-op?
Show more
the best pokemon game in the world just got f*cking Mega Evolution
PokeTide v1.41 is here, and it's 100% free
MEGA
> Mega mid-fight: hit the button, the screen goes dark, a silhouette bursts into the Mega art, and you're back in the fight
> 7 Limited Mega banners: pick which Mega you want to unlock
> Charizard X, Alakazam, Kangaskhan, Aerodactyl, Mewtwo Y, Tyranitar and Gengar, plus a shiny Megas
NEW ANIMATIONS
> crits and super-effective hits flash right on the target
> fire, grass and water hits each land with their own effect
> thunderbolt drops zigzag lightning, Dark Pulse rolls out a dark wave
> new beams, a new poof when enemies go down, fireworks after every boss
> level-up, evolution, pokeball throws and shiny sparkles all redone
> hit effects scale with your device
runs free in your browser and on your phone
which Mega are you picking?
Show more
the best pokemon game in the world just got f*cking co-op
PokeTide v1.42 is here, and it's 100% free
MEGA
> 21 new Mega Evolutions: Rayquaza, Primal Groudon, Latias, Latios, Lucario, Diancie, Absol, Gardevoir, Gallade and more
> they evolve right in the middle of a fight, the sprite changes and you keep hitting
> new Mega Mode: your run starts already Mega
> Mega Forms collection, sorted by region
CO-OP
> closed beta is live: you and a friend on one map
> clear the waves together, then take the boss down together
> testing it with the first players right now, more of you soon
DUEL
> matchmaking 1v1: you and a random player on the same seed. who's stronger?
> win duels, earn MMR, climb
runs free in your browser and on your phone
which Mega are you taking into co-op?
Show more
the best pokemon game in the world just got f*cking Mega Evolution
PokeTide v1.41 is here, and it's 100% free
MEGA
> Mega mid-fight: hit the button, the screen goes dark, a silhouette bursts into the Mega art, and you're back in the fight
> 7 Limited Mega banners: pick which Mega you want to unlock
> Charizard X, Alakazam, Kangaskhan, Aerodactyl, Mewtwo Y, Tyranitar and Gengar, plus a shiny Megas
NEW ANIMATIONS
> crits and super-effective hits flash right on the target
> fire, grass and water hits each land with their own effect
> thunderbolt drops zigzag lightning, Dark Pulse rolls out a dark wave
> new beams, a new poof when enemies go down, fireworks after every boss
> level-up, evolution, pokeball throws and shiny sparkles all redone
> hit effects scale with your device
runs free in your browser and on your phone
which Mega are you picking?
Show more
your Claude Code is f*cking burning your subscription
Anthropic just dropped a full guide on picking Fable 5.2, Opus 5.5, Sonnet or Haiku, here's it
prob: the built-in Explore subagent no longer runs on Haiku, it runs on your main model
on Opus 5.5 or Fable 5.2, every file read and every code search eats the expensive model
fix: make your own subagent called Explore with model: haiku
or just let Fable 5.2 audit your whole setup against the guide
paste this:
1. Read this guide: (link in replies)
Then audit my Claude Code setup against it: user and project settings, subagents in ~/.claude/agents and .claude/agents, and CLAUDE.md.
2. Find three kinds of problems.
2.1. Subagents on the wrong model.
Find subagents that only search files, read logs or run tests but have no model set, so they run on my main model.
If I have no custom Explore subagent with model: haiku, count that as a problem too.
2.2. Check the reverse as well: subagents that edit code or make decisions should stay on the main model.
Make sure CLAUDE_CODE_SUBAGENT_MODEL doesn't push them onto a weaker one. Check only that variable: in the env block of my settings and in my shell environment.
3. Model or effort doesn't fit the task.
Use the guide's table:
- Opus 5.5 at medium effort for everyday engineering and debugging
- Sonnet 5.5 for well-scoped work, like a bug fix or a feature from a spec
- Fable 5.2 for major refactors, migrations and leading many subagents
Flag a default effort above high and suggest a stronger model at lower effort first.
Flag opusplan as the default: it puts code edits on Sonnet.
4. No way to verify work in the project.
Check if CLAUDE.md says how to run tests and the build, and if any check exists at all.
The guide's order: make the work verifiable first, then raise effort one level, then move to a bigger model.
For every problem:
- quote the setting and the line from the guide
- explain what it costs and when it breaks
- give the smallest fix as config I can copy
5. Read the project code and recent commits.
Make a table: which kinds of tasks in this project go on which model and effort.
Finally, explain how to use /model and /usage to run one task from my list on two models and compare the cost.
Show recommendations first. Don't change anything yet.
my pokemon now explore on haiku and still find every bug
Show more
your Claude Code is f*cking burning your subscription
Anthropic just dropped a full guide on picking Fable 5.2, Opus 5.5, Sonnet or Haiku, here's it
prob: the built-in Explore subagent no longer runs on Haiku, it runs on your main model
on Opus 5.5 or Fable 5.2, every file read and every code search eats the expensive model
fix: make your own subagent called Explore with model: haiku
or just let Fable 5.2 audit your whole setup against the guide
paste this:
1. Read this guide: (link in replies)
Then audit my Claude Code setup against it: user and project settings, subagents in ~/.claude/agents and .claude/agents, and CLAUDE.md.
2. Find three kinds of problems.
2.1. Subagents on the wrong model.
Find subagents that only search files, read logs or run tests but have no model set, so they run on my main model.
If I have no custom Explore subagent with model: haiku, count that as a problem too.
2.2. Check the reverse as well: subagents that edit code or make decisions should stay on the main model.
Make sure CLAUDE_CODE_SUBAGENT_MODEL doesn't push them onto a weaker one. Check only that variable: in the env block of my settings and in my shell environment.
3. Model or effort doesn't fit the task.
Use the guide's table:
- Opus 5.5 at medium effort for everyday engineering and debugging
- Sonnet 5.5 for well-scoped work, like a bug fix or a feature from a spec
- Fable 5.2 for major refactors, migrations and leading many subagents
Flag a default effort above high and suggest a stronger model at lower effort first.
Flag opusplan as the default: it puts code edits on Sonnet.
4. No way to verify work in the project.
Check if CLAUDE.md says how to run tests and the build, and if any check exists at all.
The guide's order: make the work verifiable first, then raise effort one level, then move to a bigger model.
For every problem:
- quote the setting and the line from the guide
- explain what it costs and when it breaks
- give the smallest fix as config I can copy
5. Read the project code and recent commits.
Make a table: which kinds of tasks in this project go on which model and effort.
Finally, explain how to use /model and /usage to run one task from my list on two models and compare the cost.
Show recommendations first. Don't change anything yet.
my pokemon now explore on haiku and still find every bug
Show more
your coffee routine is f*cking robot training data now
Memories Upload just launched: you wear a tiny camera, do real tasks, get paid, and robotics labs train on your footage
HOW IT WORKS
> clip the LUCI Pin to your wrist, chest or cap
> claim a task in the app, like making a coffee
> do it normally. footage uploads itself, their AI scores it, you get rewarded
WHAT LABS GET
> every clip indexed into hand pose, head pose, object mesh, action segments and dense captions
> 3D hand pose for both hands, 21 joints, every frame
> video + annotations straight into training through their API
THE SCALE
> 1M+ hours of first-person video in 2 months
> 5,000+ pins recording worldwide
> 3 of the top 5 frontier robotics labs already train on it, per the founder
robots are learning to make coffee by watching people make coffee
my pokemon crew already claimed the coffee task..
Show more
Today we launch #
MemoriesUpload#.
Robots need internet-scale first-person video of how people get things done. Collecting this at high quality takes hardware, model and infra working together.
That's what we've spent over a year building. And already, 3 of the top 5 frontier robotics labs are training models with the data.
Show more
your coffee routine is f*cking robot training data now
Memories Upload just launched: you wear a tiny camera, do real tasks, get paid, and robotics labs train on your footage
HOW IT WORKS
> clip the LUCI Pin to your wrist, chest or cap
> claim a task in the app, like making a coffee
> do it normally. footage uploads itself, their AI scores it, you get rewarded
WHAT LABS GET
> every clip indexed into hand pose, head pose, object mesh, action segments and dense captions
> 3D hand pose for both hands, 21 joints, every frame
> video + annotations straight into training through their API
THE SCALE
> 1M+ hours of first-person video in 2 months
> 5,000+ pins recording worldwide
> 3 of the top 5 frontier robotics labs already train on it, per the founder
robots are learning to make coffee by watching people make coffee
my pokemon crew already claimed the coffee task..
Show more
Today we launch #
MemoriesUpload#.
Robots need internet-scale first-person video of how people get things done. Collecting this at high quality takes hardware, model and infra working together.
That's what we've spent over a year building. And already, 3 of the top 5 frontier robotics labs are training models with the data.
Show more
holy f*ck, someone cut the "no" out of an AI
a free 9B model that refuses 3 prompts out of 100, and it runs on your own PC
DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED, open on Hugging Face:
WHAT THEY DID
> took Qwen3.5-9B
> fed it 250K math and science examples written by DeepSeek-V4-Pro
> then cut the refusals out of it
THE RESULT
> author's test: the original said no 100 times out of 100
> this one said no 3 times
> same small 9B brain, way fewer "I can't help with that"
RUN IT YOURSELF
> the Q4 version is about 5.6 GB
> small enough for a normal PC. no API, no cloud, nobody in the middle
> Apache-2.0, 100% free
my pokemon crew already asked it where the shinies hide xd
Show more
holy f*ck, someone cut the "no" out of an AI
a free 9B model that refuses 3 prompts out of 100, and it runs on your own PC
DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED, open on Hugging Face:
WHAT THEY DID
> took Qwen3.5-9B
> fed it 250K math and science examples written by DeepSeek-V4-Pro
> then cut the refusals out of it
THE RESULT
> author's test: the original said no 100 times out of 100
> this one said no 3 times
> same small 9B brain, way fewer "I can't help with that"
RUN IT YOURSELF
> the Q4 version is about 5.6 GB
> small enough for a normal PC. no API, no cloud, nobody in the middle
> Apache-2.0, 100% free
my pokemon crew already asked it where the shinies hide xd
Show more
cloud AI is f*cking cooked
this open-source package runs AI that hears you, sees images and calls your code.
> talk to it in seven languages and get every word back with timestamps
> show it a picture and it tells you what's in it
> feed it an image and get clean structured data back
> hand it your python functions and it decides when to call them
> the models run on your machine, not on someone else's server
> want it better at your thing? fine-tune it on your own data
> Apache-2.0. take it and build
my pokemon already want ears and eyes xd
Show more
cloud AI is f*cking cooked
this open-source package runs AI that hears you, sees images and calls your code.
> talk to it in seven languages and get every word back with timestamps
> show it a picture and it tells you what's in it
> feed it an image and get clean structured data back
> hand it your python functions and it decides when to call them
> the models run on your machine, not on someone else's server
> want it better at your thing? fine-tune it on your own data
> Apache-2.0. take it and build
my pokemon already want ears and eyes xd
Show more
T3 Code is f*cking insane
open source, and it turns your AI coding agents into one team
> Claude Code, Codex, Cursor, OpenCode, Grok Build and Antigravity in one window
> runs on the subs you already pay for. they sell you nothing
> subagents: one thread hands work to helpers running in parallel
> pick the model and reasoning level per thread, switch mid-thread
> one button to commit, push and open a PR
> desktop, web, and a phone app to drive the agents on your machine
> MIT. don't like something? fork it
my PokeTide's pokemon crew already moved in xd
Show more
T3 Code is f*cking insane
open source, and it turns your AI coding agents into one team
> Claude Code, Codex, Cursor, OpenCode, Grok Build and Antigravity in one window
> runs on the subs you already pay for. they sell you nothing
> subagents: one thread hands work to helpers running in parallel
> pick the model and reasoning level per thread, switch mid-thread
> one button to commit, push and open a PR
> desktop, web, and a phone app to drive the agents on your machine
> MIT. don't like something? fork it
my PokeTide's pokemon crew already moved in xd
Show more
the best pokemon game in the world just got f*cking Mega Evolution
PokeTide v1.41 is here, and it's 100% free
MEGA
> Mega mid-fight: hit the button, the screen goes dark, a silhouette bursts into the Mega art, and you're back in the fight
> 7 Limited Mega banners: pick which Mega you want to unlock
> Charizard X, Alakazam, Kangaskhan, Aerodactyl, Mewtwo Y, Tyranitar and Gengar, plus a shiny Megas
NEW ANIMATIONS
> crits and super-effective hits flash right on the target
> fire, grass and water hits each land with their own effect
> thunderbolt drops zigzag lightning, Dark Pulse rolls out a dark wave
> new beams, a new poof when enemies go down, fireworks after every boss
> level-up, evolution, pokeball throws and shiny sparkles all redone
> hit effects scale with your device
runs free in your browser and on your phone
which Mega are you picking?
Show more
the best pokemon game in the world just got f*cking Mega Evolution
PokeTide v1.41 is here, and it's 100% free
MEGA
> Mega mid-fight: hit the button, the screen goes dark, a silhouette bursts into the Mega art, and you're back in the fight
> 7 Limited Mega banners: pick which Mega you want to unlock
> Charizard X, Alakazam, Kangaskhan, Aerodactyl, Mewtwo Y, Tyranitar and Gengar, plus a shiny Megas
NEW ANIMATIONS
> crits and super-effective hits flash right on the target
> fire, grass and water hits each land with their own effect
> thunderbolt drops zigzag lightning, Dark Pulse rolls out a dark wave
> new beams, a new poof when enemies go down, fireworks after every boss
> level-up, evolution, pokeball throws and shiny sparkles all redone
> hit effects scale with your device
runs free in your browser and on your phone
which Mega are you picking?
Show more
this is f*cking dangerous
point it at any app, site or game and your AI agent builds your own version of it
Craft Anything, open source, pure Rust:
> study -> build -> check: it learns how the target works, builds it, then shows what doesn't match
> clean room by default: it never reads the original's code, only what you can see from outside. deep mode looks inside
> runs on your own Claude Code login. Codex, Copilot and others through ACP
> 320 tools, each one a CLI command and an MCP tool, so any agent can drive it
still early alpha, and it already feels unfair
my PokeTide bros tested it on my own game first xd
Show more
the real magic of mornings is checking my polymarket positions
today especially good: everything green
until recently i was entering positions without much thought. lost money often as a result.
last month i started taking it more seriously: plugged claude cowork into the audit step.
it started clicking.
now all my ref income goes into new entries.
first small goal is $10,000 pnl, starting from zero, by year end or by summer xd
track me on my public profile:
best time to start was yesterday. second best time is today.
Show more
the real magic of mornings is checking my polymarket positions
today especially good: everything green
until recently i was entering positions without much thought. lost money often as a result.
last month i started taking it more seriously: plugged claude cowork into the audit step.
it started clicking.
now all my ref income goes into new entries.
first small goal is $10,000 pnl, starting from zero, by year end or by summer xd
track me on my public profile:
best time to start was yesterday. second best time is today.
Show more
you're running /compact every 10 minutes to fix a problem that disappears the second you give Claude one job instead of five.
Anthropic shipped a clean 4-minute video on context management.
/compact when the window fills up.
/clear to reset.
small files over big files. the practical tips work.
the video covers the cleanup. it doesn't cover why your window keeps filling up in the first place, and the answer is usually that you're holding too many jobs in one prompt.
look at what most people put in a single Claude call:
> guess the right answer
> choose between options
> set the position size
> watch the exit window
> read the news for sentiment
five tools' worth of state in one conversation. of course the window fills.
i ran into this on a Polymarket bot: the prediction-layer version ate context constantly.
/compact every 8-10 turns. i thought i needed better memory hygiene. i needed fewer jobs per call.
i stripped the prompt down to one job: veto an entry i had already decided to take.
the context window stopped filling. /compact went from every 8 turns to once a session, if ever.
over 90 days that prompt blocked 128 trades.
watch the 4 min, then ask why your window keeps needing it.
clean context. lean prompt. nothing to compact.
Show more