Coocka

coocka
building AI stuff & dropping memes
sometimes serious, mostly not

[Coocka @coocka_]

MOST OF MY AGENT'S BUDGET WAS GOING TO QUESTIONS, NOT WORK.

Traced every call for a week. 60%+ were decisions — not generation.

→ "should I retry this"
→ "which worker goes next"
→ "is this result good enough"
All of it routed through the same model that writes my code. Paying reasoning prices for a coin flip.
Put Jev in front of those calls instead.

→ Jev decides
→ GPT-5 only runs on the calls that actually need it
→ Everything else resolves in milliseconds

Tested it in shadow mode for a few days before trusting it live. Logs matched what I would've decided manually, every time.

Flipped it active. Haven't looked back.
→ Full Jev deep dive ↓

[codila @0xCodila]

Jev is the "Internet" moment for the AI industry

It tells your agents and LLMs what to do next, in milliseconds and at almost zero cost

If you set it up correctly, you will have the AI engineer’s stack for 2028

In this article, I show you how https://t.co/x3qn41ejnr

[Coocka @coocka_]

ASTRA WON EVERY BENCHMARK IN SEPTEMBER. THEN I CHECKED WHAT HAPPENED AFTER TOKEN ONE.

for two years the story was that price tracked quality. pay more, get faster, get smarter, all three numbers move together. gpt-6 astra and fable 5.1 just broke that story in public, and almost nobody read the part that actually matters.

both labs raised the same week. $10/$50. kimi k3 didn't move.

everyone turned that into a benchmark debate. it isn't one. it's a latency story, and the dashboard is the one telling it wrong.

astra's own status page reports a mean response time under 9 seconds. that number is accurate. it's also the wrong number to build a routing decision on — an average is what's left after the slow calls get diluted by everything that wasn't the problem.

i stopped trusting it after the third night my queue backed up on a metric that said everything was fine.

p50 is 6.1 seconds, nobody notices. p90 is 24 seconds, you'd notice if you were watching. p99 is 322 seconds, one in a hundred calls, every single shift. p99.9 is 481 seconds, straight into timeout territory.

mean: 9s. p99: 322s. same system. one of those numbers you put in a slide, the other one is what your users actually sat through.

at low volume, 1-in-100 sounds like an edge case. it isn't one once an agent stack is routing, scoring, and chaining tool calls hundreds of times an hour. that 1% shows up daily, sometimes hourly, and it never shows up in the number anyone screenshots for a keynote.

four things were driving the tail, and none of them were the model being "slow":

-> reasoning-depth variance — astra decides per call how hard to think, nothing caps the worst case
-> cold context reload — no cache hit, the full context gets re-read from zero every time
-> tool-call chaining — one slow tool in the chain blocks the entire response behind it
-> no typed short-circuit — a yes/no decision still gets routed through a full text generation pass
fix all four and the tail doesn't disappear. it moves. that's the part nobody selling the $200 seat wants said out loud.
astra wins the benchmark, and it's not close. run the same 40 tasks and it scores higher on almost every axis that shows up on a chart.
kimi k3 wins the bill, by more than the benchmark gap justifies. 84% cheaper per token isn't the headline. the headline is that the quality gap is small and the tail-latency gap is not, and the tail is the part that actually reaches your users, not the part that reaches a chart.

while both labs were being compared on sticker price, fable 5.1 quietly undercut both of them on cache. same task, same prompt structure, run through cache and the real bill drops by more than the headline number implies. nobody covered this because it doesn't fit in a price-comparison table.

nothing dramatic changed on my end. one routing rule.

classify and score calls go to a typed, deterministic path. generation-heavy calls stay on astra. anything cache-eligible gets checked before it touches a model at all.

the account that used to eat a p99 spike four times a night now eats it once. the bill dropped. the tail didn't vanish, it just stopped being my problem every hour.

fast average, honest tail — pick a model whose keynote number and dashboard number are close to the same thing. right now none of the three headline models earn that in one sentence. the closest one changes depending on which call type you're routing.

the full routing config, the 322-second logs, and the cache math that got this stack out of the tail — in the piece below ↓

[Coocka @coocka_]

1,092 BROWSER CALLS. THEN JEV CUT IT TO 101.

Same task. Same browser. 91% fewer calls.

→ the model doesn't see the whole page anymore — just a short, fresh action table
→ pick operation + target, that's the only decision it makes per step
→ the small LLM only wakes up when a field actually needs typing
25% lower median time. Not from a faster model. From asking it fewer questions.
Zürich → London, 7.1 seconds, real time, no speed-up. Official verified run.

→ Full Jev deep dive ↓

[Coocka @coocka_]

@Solerfin No lie, read it twice 🤝

[Coocka @coocka_]

cold context was the worse offender by far. chaining adds latency but it's predictable — you can budget for it. cold context spikes are what wreck your p99 because they're rare enough to not show up in dashboards until someone's staring at a 8s outlier wondering why the mean looks fine.

[Coocka @coocka_]

ASTRA WON EVERY BENCHMARK IN SEPTEMBER. THEN I CHECKED WHAT HAPPENED AFTER TOKEN ONE.

for two years the story was that price tracked quality. pay more, get faster, get smarter, all three numbers move together. gpt-6 astra and fable 5.1 just broke that story in public, and almost nobody read the part that actually matters.

both labs raised the same week. $10/$50. kimi k3 didn't move.

everyone turned that into a benchmark debate. it isn't one. it's a latency story, and the dashboard is the one telling it wrong.

astra's own status page reports a mean response time under 9 seconds. that number is accurate. it's also the wrong number to build a routing decision on — an average is what's left after the slow calls get diluted by everything that wasn't the problem.

i stopped trusting it after the third night my queue backed up on a metric that said everything was fine.

p50 is 6.1 seconds, nobody notices. p90 is 24 seconds, you'd notice if you were watching. p99 is 322 seconds, one in a hundred calls, every single shift. p99.9 is 481 seconds, straight into timeout territory.

mean: 9s. p99: 322s. same system. one of those numbers you put in a slide, the other one is what your users actually sat through.

at low volume, 1-in-100 sounds like an edge case. it isn't one once an agent stack is routing, scoring, and chaining tool calls hundreds of times an hour. that 1% shows up daily, sometimes hourly, and it never shows up in the number anyone screenshots for a keynote.

four things were driving the tail, and none of them were the model being "slow":

-> reasoning-depth variance — astra decides per call how hard to think, nothing caps the worst case
-> cold context reload — no cache hit, the full context gets re-read from zero every time
-> tool-call chaining — one slow tool in the chain blocks the entire response behind it
-> no typed short-circuit — a yes/no decision still gets routed through a full text generation pass
fix all four and the tail doesn't disappear. it moves. that's the part nobody selling the $200 seat wants said out loud.
astra wins the benchmark, and it's not close. run the same 40 tasks and it scores higher on almost every axis that shows up on a chart.
kimi k3 wins the bill, by more than the benchmark gap justifies. 84% cheaper per token isn't the headline. the headline is that the quality gap is small and the tail-latency gap is not, and the tail is the part that actually reaches your users, not the part that reaches a chart.

while both labs were being compared on sticker price, fable 5.1 quietly undercut both of them on cache. same task, same prompt structure, run through cache and the real bill drops by more than the headline number implies. nobody covered this because it doesn't fit in a price-comparison table.

nothing dramatic changed on my end. one routing rule.

classify and score calls go to a typed, deterministic path. generation-heavy calls stay on astra. anything cache-eligible gets checked before it touches a model at all.

the account that used to eat a p99 spike four times a night now eats it once. the bill dropped. the tail didn't vanish, it just stopped being my problem every hour.

fast average, honest tail — pick a model whose keynote number and dashboard number are close to the same thing. right now none of the three headline models earn that in one sentence. the closest one changes depending on which call type you're routing.

the full routing config, the 322-second logs, and the cache math that got this stack out of the tail — in the piece below ↓

[starmex @starmexxx]

https://t.co/ErEYcWPWBo

[Coocka @coocka_]

1,092 BROWSER CALLS. THEN JEV CUT IT TO 101.

Same task. Same browser. 91% fewer calls.

→ the model doesn't see the whole page anymore — just a short, fresh action table
→ pick operation + target, that's the only decision it makes per step
→ the small LLM only wakes up when a field actually needs typing
25% lower median time. Not from a faster model. From asking it fewer questions.
Zürich → London, 7.1 seconds, real time, no speed-up. Official verified run.

→ Full Jev deep dive ↓

[codila @0xCodila]

Jev is the "Internet" moment for the AI industry

It tells your agents and LLMs what to do next, in milliseconds and at almost zero cost

If you set it up correctly, you will have the AI engineer’s stack for 2028

In this article, I show you how https://t.co/x3qn41ejnr

[Coocka @coocka_]

@Solerfin saving this

[Coocka @coocka_]

@1onetouch1 solid breakdown ngl

[Coocka @coocka_]

@zukooxbt wait this is genuinely useful, more of this pls

[Rozetka @rozetka_ua]

Шановні Відьмаки, приготуватися — даруємо The Witcher 3 в Steam. Буде 3 переможці

Щоб ви змогли безплатно ремастер отримати та НАРЕШТІ українську локалізацію зацінити, треба:

- лайкнуть
- ретвітнуть
- на нас підписатися

Аард та Ігні лишаємо при собі, а результати будуть 15 вересня 😉

[Coocka @coocka_]

@SynthPotato @IndieGameJoe PS - the only choose!

[Miilato @litteralyme0]

love you bro

[Rozetka @rozetka_ua]

Чекаєте GTA VI? А ми її подаруємо вам ;)

Один щасливець отримає коробку з кодом активації для PS5, щойно вона зʼявиться в Україні 😏

1. Лайк
2. Репост
3. Підписка

4 серпня рандомно визначимо переможця. Нехай щастить!

[Coocka @coocka_]

@Heistedxyz 5gVtFR21jiSa4iZsxaUEEHyBBnmLoExfNGYynkvPD24p

[Coocka @coocka_]

A GUY SHOWED HOW HIS $200 CLAUDE BILL JUST DISAPPEARED, REPLACED BY A PRIVATE AGENT NAMED HERMES THAT LIVES WHEREVER HE ALREADY CHATS

not a tab you have to remember to open, an agent that becomes an actual contact in the app you use

at 0:09 you can see the setup screen, pick your channel, telegram ready today, just the web with no phone needed, whatsapp and imessage both listed as coming soon, so it's built to live across all of them, not locked to one app

at 0:18 it's literally "hatching", learning his name, his tone, his goals before it sends a single message, onboarded more like a hire than installed like software

once it's live it just sits in his messages like any other contact, always on, no $200/month subscription wall behind it

the shift isn't the model, it's the delivery, AI stops being a tab you visit and becomes a contact that's just always there

https://t.co/uW9eKaECFl

worth watching how fast "my agent's in my texts" replaces "let me open claude" as the default move

[Coocka @coocka_]

APPLE IS ABOUT TO HAND THE "MAC MINI AI FARM" CROWD A MASSIVE UPGRADE, AND MOST PEOPLE STILL THINK IT'S JUST A DESKTOP REFRESH

multiple leaks now point to the new mac mini launching with the M5 chip, both a standard M5 and an M5 Pro variant, base model finally starting at 16GB of ram instead of the 8GB people have been complaining about for years

leaked benchmarks put cpu speed 25-30% faster than the M4, but the real story is the neural engine, reportedly redesigned for roughly 4x the local AI performance

that's the part that actually matters for everyone running ollama, vllm, or local agents off a mac mini closet setup, this isn't a spec bump, it's apple quietly building the chip the AI-rig crowd has been hacking together with used hardware

WWDC timeline points to a june unveil, store shelves globally by july

every kenji, chen and linh running a private inference box off a $599 mac mini is about to get a free performance upgrade just by buying the next gen

apple isn't trying to compete with nvidia, they're just making sure the cheapest entry point into local AI gets harder to ignore

worth watching how fast the "mac mini AI farm" niche scales once the hardware catches up to what people are already doing with it

#AppleM5 #MacMini #AppleIntelligence #EdgeAI #LocalAI #AIHardware #TechNews #AIagents

[Coocka @coocka_]

@Crypto__Haris 5gVtFR21jiSa4iZsxaUEEHyBBnmLoExfNGYynkvPD24p

[Coocka @coocka_]

@pandaa 5gVtFR21jiSa4iZsxaUEEHyBBnmLoExfNGYynkvPD24p