Haroon

20 | ai/backend dev @ | building agents, breaking systems

[ajay yadav @BetterSayAJ]

My current take is that people often lump together two pretty different ideas:

• Model routing = pick the best model for the task
• Model councils = ask multiple models, then aggregate
The goals are different.

Model councils are mostly about pushing performance/frontier capability. We've seen this recently with OpenRouter Fusion, Sakana's Fugu, and a bunch of work around high-agency systems. Makes a lot of sense as a tool for hard problems where extra compute is worth it.

Routing is different. To me, the main reason to do routing is cost.

I'm probably a bit more bullish on routing than Dax though. Expecting most users to know exactly which model to use for which task doesn't seem realistic outside of power-user niches. The tricky part is things like prompt caching and context continuity. My intuition is that once you've routed a conversation, you should be heavily biased toward staying on that model unless there's a strong reason to switch.

That said, routing isn't the first lever I'd pull.

There are simpler ways to control spend: model-specific budgets, spending caps, usage policies, etc. ("You get X dollars of Opus, then fall back to Sonnet" is almost easier to reason about than a complex router.)

[Dhravya Shah @DhravyaShah]

Memory on the harness level

after 2 years... i've started writing blogs again!

In this post, I dive _very_ deep into the different ways to compose and add memory to the harness for various different use cases.

it's interactive. it's fun. you can listen to it too!

[Command Code @CommandCodeAI]

New in v0.40.0 ⋅ /fork

Fork new sessions from current context
The original stays intact.
/resume returns you to the original.

Try different directions without breaking your sessions.

[shadcn @shadcn]

No model survives the "are you sure?" They all fold instantly.

[Kai @hqmank]

1/

Codex is quietly killing your SSD.

It writes diagnostic logs to disk non-stop, even when you're not
doing anything. Your SSD has a write limit. Codex is burning through it in the background.

One command fixes it 👇

[ℏεsam @Hesamation]

Vibe coding essentials.

[hardmaru @hardmaru]

Human intelligence is fundamentally a collective intelligence. We solve complex problems by participating in a vast cultural network that builds upon ideas across generations.

I believe the strongest AI systems will become a collective intelligence, too.

Since we started Sakana AI, our core conviction has been that the most powerful AI systems will be collaborative ecosystems, not isolated monoliths. Evolution innovates under constraints, and the future belongs to systems that explicitly learn how to coordinate collective intelligence.

Today, we are taking a major step toward that future with the launch of Sakana Fugu.

Fugu dynamically orchestrates the world’s best models to tackle complex tasks. We are proving that a well-orchestrated pool of swappable agents can match restricted frontier models like Fable and Mythos.

But Fugu is about more than just performance. I believe that Orchestration Models are the next frontier, beyond bigger models.

Relying on a single company’s model for national infrastructure is a massive risk. As recent export controls have shown, access to top models can disappear overnight.

Collective intelligence is the practical hedge against this concentration of power. Fugu simply routes around vendor restrictions by relying on an entirely swappable agent pool.

I am incredibly proud of our Tokyo team for shipping this. By orchestrating the world’s models, we are delivering the resilient blueprint required for AI sovereignty.

Read our full vision and results here:
https://t.co/EONDdWx5Ld 🐡

[Daniel @growing_daniel]

he's halfway through a life altering refactor of the entire codebase, now make the model illegal

[Reed @reed_barnes]

sorry buddy, time to go

[Theo - t3.gg @theo]

Want a good "first loop" to use with Codex?

"When you are done designing the API, get a second opinion from Opus with 'claude -p'"

This has significantly improved the quality of the code I get out of OpenAI models.

[Kent C. Dodds 🏹 @kentcdodds]

If GPT 5.6 still requires me to say "no fallbacks please" I think I'll crash out.

[Samay @Samaytwt]

DocuSign is basically a document-signing app.

So why does a document signer need 6,838 employees?

[Zhu Liang @paradite_]

i’m really surprised that people don’t see this.

It’s mathematically true that llms can’t come up with novel ideas, because the whole point of training is to reduce loss, gain rewards so that the model adhere to rules and ground truth.

if you have a model that can come up with novel ideas, it must have high loss during sft or rl.

[atharva ☆ @k7agar]

this website is so entertaining

[Jay Hack @mathemagic1an]

ChatGPT launched November 2022

Startups as a category have since become more concentrated in capital/talent

This will hit a new equilibrium w/ big tech but we have not yet hit the floor. The reign of the 2-man YC co is at a close

[NIK @ns123abc]

“Sir… John Jumper… the director of DeepMind… the co-creator of AlphaFold… the man who won the Nobel Prize with you… sir… he just announced he’s leaving Google DeepMind and joining Anthropic…”

[Thorsten Ball @thorstenball]

How to get better at prompting

[Haroon @skywalkerr0x]

@al0kdangre Are you doing freelancing

[Haroon @skywalkerr0x]

For 1$ this is my usage in @CommandCodeAI