Eddy Goh

Tech direction that makes products succeed. CTO of 3 companies. I write about building things that work.

[Eddy Goh @eddygte]

still running gpt-4.1 in production?

you're paying 8x more for less than half the intelligence.

gpt-4.1: index 19 at $3.50 blended per 1m tokens.

gpt-5.6 luna: index 51 at $0.45.

after the 80% cut, staying on the old model is the expensive choice.

[Eddy Goh @eddygte]

@CricTalk29 u don't

[Eddy Goh @eddygte]

poker first, but it handles blackjack, big two, mahjong and monopoly too.

free on ios and android → https://t.co/eQlsz4FGPo

how does your home game handle the cash-out at the end of the night?

[Eddy Goh @eddygte]

one player creates a room. everyone else scans a qr code and they're in.

no accounts, no signups. under 30 seconds from download to first hand.

[Eddy Goh @eddygte]

the interesting engineering problem was settlement.

after a long session, 8 players owe each other in tangled directions. we solve for the minimum number of transfers to clear it all.

one tap. debts gone.

[Eddy Goh @eddygte]

every home poker game ends the same way. someone miscounts a stack, and ten minutes go to arguing over who owes who.

so we built partypot. a digital banker for poker night.

no chips, no cash, no ious.

[Eddy Goh @eddygte]

Next time OpenAI or Moonshot releases something, don't ask if Anthropic can catch up.

Ask if they want to release it right away or take a sip of coffee first...

[Eddy Goh @eddygte]

Anthropic released Opus 5 yesterday.

Half the price of Fable 5, and it scores higher on most tests.

That's 8 days after Kimi K3.

15 days after GPT-5.6 Sol.

Nobody builds a model this good in 8 days.

[Eddy Goh @eddygte]

Every AI lab uses its newest unreleased model to help build the next one.

So what you can buy is never their best.

It's their best minus a head start they keep for themselves.

My guess: 4 to 8 months.

[Eddy Goh @eddygte]

Fable officially included in Claude normal subscription plan!

Thanks to the competitors ▓▓▓

[Eddy Goh @eddygte]

scale isn't a framework you install.

it's a hundred boring decisions. cache, cdn, static, queues.

what's the first thing that breaks in your stack under load?

[Eddy Goh @eddygte]

1 extra join = 1.1 million extra queries at our scale.

at high load, a structured app won't auto-optimize. and it's not about indexes.

redis cache, lazy load, cdn, static builds. you earn every millisecond by hand.

[Eddy Goh @eddygte]

so: keep reads static. push writes to a queue.

every transaction gets queued and processed by microservices.

now the database isn't the bottleneck for reads or writes.

[Eddy Goh @eddygte]

Also when both hardware and LLM advancement met, there's a point that AI will live on merely our TV, mobile phone, CCTV camera, no longer need a large data center to do all* the important work.

[Eddy Goh @eddygte]

Kimi K3 is "better" than Fable 5 & GPT-5.6 Sol?

The competition getting hotter where the whole world are advancing to create a better model.

Kimi K3 from China even release it as open source model.

[Eddy Goh @eddygte]

Eventhough it might be lesser compared to the enterprise lead model like Fable 5.

I believe at most 2 more years, you can use these open source model for your most important work.