clay @deforestpeg

I build AI agents and honest backtesting tools, and post the real results. github.com/claygeo Joined May 2022

Tweets

121
Followers

1K
Following

2K
Likes

132K

clay @deforestpeg

12 hours ago

Watch live: codexplays.games/pokemon-red

0 0 2 77 0

View Details

Last update ended with the agent taking the Poke Flute from Pokemon Tower. This is why. The Snorlax blocking Route 12. Its first move after the rescue: walk up to the sleeping roadblock, open the bag, play the flute. Took the fight, cleared the road, kept moving south. Nobody coded that in. The model just knows Pokemon.

2 1 8 614 0

View Details

clay @deforestpeg

15 hours ago

ive run agents on hard caps from day one, flat subscription, nothing to fall back on. the cap is what taught me which steps actually need a model and which were just expensive habit. spend limits dont end the token maxxing era, they end the part where nobody knew what the tokens were buying

0 0 3 142 0

View Details

clay @deforestpeg

15 hours ago

running agents on flat budgets forced me to measure exactly this, tokens to done is the only number that ends up mattering. the model that wins the one shot benchmark can quietly lose in a loop once retries and wandering count against it. accuracy is the sticker price cost to goal is what you pay

0 0 1 99 0

View Details

clay @deforestpeg

a day ago

watch it think in real time: codexplays.games/pokemon-red

0 0 1 244 0

View Details

clay @deforestpeg

a day ago

Badge 4 of 8. Same save, no resets. Then the wall: Pokemon Tower broke the agent for days. So I rebuilt it — Codex picks the objectives now, the machinery just walks. First night on the new brain: beat the ghost Marowak, rescued Mr. Fuji, took the Poke Flute. On its own.

2 0 9 591 1

View Details

clay @deforestpeg

a day ago

@businessbarista ive cut agent spend over 50% without touching the model, just pulling the steps that dont need one out of the loop. most agent cost is a wasteful loop, not an expensive model. finetuning a cheaper one is fixing a bill you ran up building it wrong

1 0 4 438 1

View Details

clay @deforestpeg

a day ago

@ox_vanguard my dms are always open!

1 0 1 18 0

View Details

clay @deforestpeg

3 days ago

17% fee APR on this USDC-SOL range. after impermanent loss it nets +$4 on $10k. thats the whole problem with DLMM LPing, the APR looks great and IL quietly eats it. binsight runs your exact range against real on chain price, volume + fees and shows the net.

4 1 14 767 0

View Details

clay @deforestpeg

a day ago

@ox_vanguard its passive market making, yeah. youre putting up two sided liquidity in a range and collecting swap fees, but you also eat the inventory risk an active MM would hedge out. usually the fees dont cover it, and thats the gap binsight measures

1 0 1 9 1

View Details

clay @deforestpeg

a day ago

@karinanguyen the tools already do the inspiring, more than a message could

0 0 1 237 0

View Details

clay @deforestpeg

a day ago

@alexatallah @cclark @OpenRouter from routing tokens to routing airflow

1 0 1 130 0

View Details

clay @deforestpeg

2 days ago

@ollama @NousResearch self generating skills is the right shape, but improves them as you use them is where it gets hard. the question is whether it can tell a skill actually got better or its just confidently rewriting it. self improvement is a verifier problem, not a generation one

0 0 2 213 1

View Details

clay @deforestpeg

2 days ago

@enzo_gte the design system port is the real tell, not the speed. rebuilding from the ground up instead of jamming the old language into the new scheme is the model making a structural judgment call, not pattern matching. thats the actual jump everything else is throughput

0 0 1 781 0

View Details

clay @deforestpeg

2 days ago

@_xjdr the on par with k2.6 outside the new claude code features line is the whole story imo. the meaningful delta this release is the harness, not the model. raw capability is converging, the agent tooling around it is where the gap actually is now

0 0 5 661 0

View Details

clay @deforestpeg

2 days ago

@EricNewcomer the fatigue is real but the unlock is that for most tasks the model barely matters. pick one default, use it for everything, only switch on the long hard multi step stuff. the per task optimizing costs more time than it saves

0 0 0 90 0

View Details

clay @deforestpeg

2 days ago

@ayushagarwal this is the real signal, not the demo stuff. correlating pod logs, sql errors and image digests to catch a cached image hitting a migrated db is exactly the cross system glue models usually faceplant on. if its doing that consistently, thats the actual jump

0 0 2 346 0

View Details

clay @deforestpeg

2 days ago

@jayelmnop the capability jump looks real, no question. but running out of hard tasks is the line id push, the hard ones left arent benchmarks, theyre the boring brittle stuff, build systems and long horizon agentic drift. it keeps acing the demo and tripping on the glue

0 0 2 118 0

View Details

clay @deforestpeg

2 days ago

@jakehalloran1 every model is the smartest ever for the first 2 hours. claude psychosis has a half life, it ends the moment it confidently breaks your build

1 0 90 9K 2

View Details

clay @deforestpeg

2 days ago

@M1Astra three checkpoints in testing at once is the real signal here. the cadence is now faster than anyone can actually adopt a model, let alone benchmark one

0 0 4 1K 0

View Details

clay @deforestpeg

2 days ago

@Rex_Ogjy exactly. and high APR usually means high volatility, which is the same thing feeding the IL, so the juiciest looking pools are often the worst net. the APR almost warns you