Haya

Red queening AI, enjoying dad life

[Haya @HayagrivS]

@adityaag And soon agents will also go from needing CPU threads to multi-core CPUs to Kubernetes clusters to do their work

[Haya @HayagrivS]

@mattshumer_ @bot Inevitable that every Bot will become a Kubernetes cluster

[Haya @HayagrivS]

@joshearle One reason I can think of is Monolithic vs Services. As your agents start to scale and do 1000s of tasks, one agent becomes a choke point.

[Haya @HayagrivS]

Ok REAL TALK.. Grok Bot is legit. The greatest products take infrastructure complexity and abstract it away. This is an iPhone "It just works" moment. Wow! @cursor_ai @spacexai You guys crushed it!

[Grok Bot @bot]

Introducing Grok Bot, now in early beta.

Bots are AI teammates that do real work for you. They sign in to your tools, use them just like you do, and come back with finished work.

[Haya @HayagrivS]

Boss, I get you are a rockstar. But since you ask us how you can make Codex better maybe automate every chance you get for a reset instead of doling them out during your p*** contests. I am saying this as an almost exclusive Codex user and lover.

[Tibo @thsottiaux]

That's right, GPT-5.6 Sol is awesome and can be used pretty much anywhere, including in the CC harness.

To celebrate this, together with the fact that I'm not going anywhere... I have reset usage limits for all paid users of ChatGPT Work and Codex.

Have fun out there!

[Haya @HayagrivS]

@ayushjaiswal Respectfully, this is an idiotic take. Have they acquired compute? No. Have they focused on hard domains like bio and Robotics? No. Have they done enough vague posting? Hell No.

[Haya @HayagrivS]

I think all three solve specific problems -

1. Models to scale the intelligence - intelligence density per token
2. Co-training to get you better on harness-specific tool calling - computer use, or inbuilt Gmail connector for example
3. Independent value comes from use case specific expressions - company specific harnesses come to mind

The tradeoffs are complexity of the problem vs efficiency of tool use vs LLMs cannot read your mind

[Haya @HayagrivS]

The next big job title is going to be Knowledge Development Engineer. Job description - Write workflows into environments that AI can hillclimb on.

[Haya @HayagrivS]

Oh No! Did they use distillation?

[Claude @claudeai]

Introducing Claude Opus 5.

It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

[Haya @HayagrivS]

But don’t you run into the ergodicity issue? If I don’t participate in the exponential then I might be obsolete?

[Haya @HayagrivS]

ChatGPT Voice mode is something else.... damn...

[Haya @HayagrivS]

Deterministic software in a Fable world is like a Super Mario Flag. It's a checkpoint in your journey to beat the final boss.

[Haya @HayagrivS]

@miramurati @tinkerapi @miramurati what about Fable 5 and using something like dynamic workflows for this task? One of the biggest hesitations I have towards fine-tuning is we are still on an exponential and the jumps are not always obvious version to version.

[Haya @HayagrivS]

@pitdesi Umm.. I'd have agreed for any team except the one that plays in the Bay Area.

[Haya @HayagrivS]

@snoopy_dot_jpg are you able to create the simplest verifiable loops that work without becoming slop? if yes then why is this hard to digest?

The key is Gall's law. You don't design a complex loop, you start with simple loops that work, and stitch them together.

[Haya @HayagrivS]

Severely underestimating two things IMO

1. Firms like K&E will create 100x more unique data from their use of AI than what they have already have. Differentiation should be looked at over future years, not just the past.

2. Software fluency is a non-negotiable for every firm going forward. You can wait for Harvey and Legora to make sense of every model release, and help you build through their filtered lens. Or you can develop new competencies.

AI is so empowering and exciting. Everyone wants to build now. The old world assumptions that businesses just want to solve problems and buy solutions doesn't hold. We need to stop shoving SaaS at people and be bullish on people's desire to build.

[FleetingBits @fleetingbits]

some thoughts on kirkland building its own harvey

1) kirkland is spending $500m over four years in order to build its own internal ai legal tools; kirkland intends to spend $100m this year

2) i suspect that kirkland is doing this because they have told themselves that they have valuable data and because they want to appear differentiated

3) i think the first issue is that kirkland probably does not have differentiated data from other elite law firms; at least, not at the level a harvey would absorb

4) all the elite firms probably have similar internal workflow data and so long as some of them defect, that is enough to commoditize the data kirkland wants to use for its platform

5) and, to the extent that they do have different internal workflows, harvey and legora will end up representing a better version of them and this will put kirkland at a disadvantage

6) moreover, companies like kirkland will have difficulty building their internal legal platforms because they do not have experience with software development

7) and, there are both cultural and structural issues with them managing software developers, like they cannot give non-lawyers equity in the firm due to regulation

8) so, i think firms like kirkland are better off using tools like harvey and legora and then looking to focus on where their value really is now: client relationships, local knowledge (litigation, regulation) and legal r&d (novel structures, etc...)

9) anyway, this seems to me like a phenomenon that ai creates across a lot of industries, where firms that were previously vertically integrated become unbundled due to ai because part of the intelligence gets moved to the labs or otherwise gets commoditized

10) and so, a new set of companies are created whose job it is in order to provide services complementary to the labs: forward deployed like harvey and legora and data providers like mercor, surge and handshake

[Haya @HayagrivS]

@karrisaarinen Beautiful take! Speed and velocity are completely different things.

[Haya @HayagrivS]

@sidbid Will this be available in the Claude Agent SDK as well?

[Haya @HayagrivS]

@arena @OpenAI GPT 5.5 low maybe? These results are weird and not reflective of reality.

[Haya @HayagrivS]

How do I send a DAG as a prompt to Claude? @trq212 @bcherny @AnthropicAI @claudeai

My use case is I want agent to deterministically do certain things, and then apply intelligence and choose paths.