Anunay Arunav

Cofounder, inshorts

[Anunay Arunav @anunayarunav]

Good thinking.

[spor @sporadica]

Absolutely spot on.

“Everyone will build their own personal software” (SaaS is dead!) is the biggest, most pervasive, and completely incorrect theory being thrown around right now.

Almost no one will build their own software. No matter how good AI gets.

[Anunay Arunav @anunayarunav]

Opus 4.8 might be less smart, but it’s much more collaborative than fable and opus 5 both.

Fable likes to jump ahead to solve the problem, however it understands what I am asking. Well most of the time.

Opus 5, just likes to jump. Without understanding.

Opus 4.8 however. It likes to engage. It likes to listen to you, bring trust in the conversation by suggesting the proof. And then it delivers itself. Greatest model so far for collaborative work IMO.

[Anunay Arunav @anunayarunav]

I am done with Opus 5. It just doesn’t get me.

Hope Opus 5.1 understands us better.

Switching back to 4.8. Anyone else doing the same?

[Anunay Arunav @anunayarunav]

@dotgil Yes I let fable spec and verify opus work through Claude channels api.

[Anunay Arunav @anunayarunav]

Fable is a great architect. Especially good if you challenge its assumptions and push back on its ideas.

Opus 5 excels at writing code and can keep going for days.

The best workflow is to talk to Fable before you start. Have it design the spec first, then hand it over to Opus 5 for implementation.

[Anunay Arunav @anunayarunav]

@Steve_Yegge I couldn’t agree more. I somehow feel it’s worse. It keeps going without reporting and is very overconfident in it.

[Steve Yegge @Steve_Yegge]

Welp, back to Fable. I'm done with Opus 5 as a collaborator. Opus 5 is still just Opus: blustering and often wrong.

Fable is the only enterprise-grade model on the market today. The rest of them cannot be trusted. If you're an engineer at an enterprise who's pissed off at how bad AI is, then you're right -- it is all bad. Except Fable itself, which is good enough, but too expensive to be practical. The world isn't going to have good-enough cheap-enough models for a while, it would seem.

[Anunay Arunav @anunayarunav]

Notes on the Post-AI World
(written the day Fable was released)

WE STILL UNDERESTIMATE THESE THINGS

Most people treat LLMs as chatbots — something that answers small questions and writes code. That framing hides the strangeness of what exists: systems with internal representations of nearly the entire written record, able to converse with anyone on anything. Not "knowing everything" — their knowledge is broad but uneven — but holding more usable breadth than any human ever has. And that's the small models.

Fable came out today. It's expensive; my sessions expire fast even across two accounts. But measured against the work it does, it's plausibly the cheapest high-quality labor a single person has ever been able to buy.

THE INSTANT EDGE: EXECUTION STAMINA

For people who work best solo, the binding constraint was never thinking stamina — the best entrepreneurs have infinite appetite for thought. The constraint is execution stamina: the context-switching that planning and shipping demands, which human brains handle badly and neurodivergent brains handle worse.

That's now offloadable, at least in software. Anyone who can parallelize — run many agents, keep them all productive — removes the bottleneck that kept solo builders slow. The gap between the best and the rest is far smaller than pre-AI. Fast tinkering is now the most efficient way to learn, and the advantage goes to whoever adapts fastest.

BUT PARALLELIZATION IS ONLY THE FIRST RUNG

Orchestrating agents is a must-have skill, not a moat. As the cost of intelligence falls, everyone gets fifty agents.

The next rung is what you point them at. If everyone has PhD-level capability at their fingertips, raw knowledge is commoditized — so the scarce thing is judgment: finding the most efficient solutions to the most pressing problems. That requires breadth, but breadth alone isn't enough. Judgment is knowledge plus feedback loops plus skin in the game.

Prediction: five years out, the people on top won't be hustlers. They'll be genuinely multi-disciplinary — able to see across fields well enough to know what's worth solving, and to verify agent output rather than just consume it.

THE GAP PERIOD

This assumes everyone actually has these agents at their fingertips. Not true today. Filling that gap is THIS period — and it's where a generation of entrepreneurs will be built.

It'd be a mistake to assume "agents at your fingertips" means chat forever. Look at how humans serve humans — talking is only one mode:

Media: humans transfer knowledge through artifacts others consume on their own time.
Pedagogy: humans build systems for how others learn, refined over generations. (Agents with real-world feedback may develop pedagogy faster than we ever did.)
Machines: humans build machines for others' convenience. Chat doesn't replace a machine — machines create habitual, operational learning.

Until we can rewire brains directly, we need one machine above all: something that transfers and ingrains knowledge in humans. I don't think it's chat. I think it's closer to games. The idea isn't new — gamified learning has a graveyard behind it. What's new is that generation cost collapsed, making personalized interactive learning possible for the first time. That's the unlock.

TWO EVOLUTIONS, THREE CLASSES OF BUILDERS

Human–agent interaction converges on two modes:

1. Humans asking agents to DO things — autonomous tasks, short or long.
2. Humans asking agents to BUILD things for human consumption — machines that put knowledge inside human minds.

These won't come from the same companies. The first class is already crowded. The second barely exists beyond image and video gen.

And there's a third class the first two create demand for: trust infrastructure. As delegation scales, someone has to build the evaluation, auditing, and accountability layer that lets humans trust agent output. That layer becomes “load-bearing” fast.

What a time to be alive.

[Andrej Karpathy @karpathy]

RE: "how often do you see teams actually fine tuning LLMs?"
It's an interesting question, about how prompting (optimization over prefix tokens) and finetuning (optimization over weights) will be used over time. If people have data points please pitch in.
I expect that finetuning is still quite new / a lot more involved (accessible data collection, optimization, expertise around making it work) but a lot of this is improving.

[Anunay Arunav @anunayarunav]

IAF jet lands successfully on Yamuna expressway
http://t.co/AxmEw2VDfO
-via Newsinshorts

[Anunay Arunav @anunayarunav]

Google turns images into time-lapse video
http://t.co/AxmEw2VDfO
-via Newsinshorts

[Steve Milan @stevemilan]

Still blown away by this photo.

[Sachin Bansal @_sachinbansal]

Waiting for the day when IIT campus conversations are like: "Yaar, I couldn't come up with a startup idea, so I'll have to join a job"

[Anunay Arunav @anunayarunav]

Flipkart ties up with Mumbai Dabbawalas
http://t.co/8rcZmCTiVR
-via News in Shorts

[Anunay Arunav @anunayarunav]

Big B sends hand-written note to Ayushmann
http://t.co/sD0B0wNZyq
-via News in Shorts

[George RR Martin Parody @GeorgeRRMartin_]

.@DPRK_News I'm terribly sorry if I've offended you. I'm hosting a wedding soon, perhaps you can attend? #GuestofHonor #RainsofCastamere

[Anunay Arunav @anunayarunav]

First photo from selfie stick clicked in 1926
http://t.co/u5nmZioR0f
-via News in Shorts

[Anunay Arunav @anunayarunav]

Crisp News within 60 words https://t.co/TKdW27arzB

[Anunay Arunav @anunayarunav]

Trouble for Dreamliners: This time, door won't close http://t.co/3Daw7fxgAz

[Anunay Arunav @anunayarunav]

Kahaani to be remade in Hollywood http://t.co/aw1f0wx5Qm