Prabhat Singh

Building the next generation of AI × Web2 products

[Prabhat Singh @prabhatxz]

Reasoning models are impressive on hard problems.
They still struggle when the environment is messy, the tools return partial results, and the goal itself is slightly ambiguous which is most real-world work.

[Prabhat Singh @prabhatxz]

Agent security is moving from theoretical to urgent.
Once agents can take real actions (send emails, move money, change configs), a single prompt injection or tool misuse stops being a demo fail and starts being an incident.

[Prabhat Singh @prabhatxz]

Small language models are quietly winning a lot of production use cases.
When latency, cost, and data privacy matter more than raw IQ, a well-routed SLM + good tools often beats a frontier model that has to think about everything.

[Prabhat Singh @prabhatxz]

Inference is becoming the real bottleneck, not training.
Most teams are still optimizing for the model that looks best on a leaderboard instead of the one that delivers the lowest cost per successful task in production.

[Prabhat Singh @prabhatxz]

Always-on agents are the new frontier.
The hard part isn’t giving them a goal.
It’s designing the control layer so they can run for hours without drifting, looping, or burning budget on useless actions.

[Prabhat Singh @prabhatxz]

Tool use is only half the problem.
The harder half is deciding when the model should stop, ask for clarification, or hand control back to the user instead of hallucinating a confident next step.
These are new, deeper on core AI systems issues, and ready to post

[Prabhat Singh @prabhatxz]

Everyone optimizes for the happy path.
In production AI systems the unhappy path (timeouts, partial failures, malformed tool outputs, distribution shift) is where most of the engineering time actually goes.

[Prabhat Singh @prabhatxz]

The current generation of models is surprisingly good at local reasoning and surprisingly bad at global consistency.
They can solve a hard sub-problem and then quietly contradict themselves three steps later.

[Prabhat Singh @prabhatxz]

@ManasKushwaha29 ChatGPT is currently better at tool calling and browser control, agreed
Claude still feels more reliable once the task gets long and needs consistent memory.
What kind of LinkedIn automations are you running?

[Prabhat Singh @prabhatxz]

Fine-tuning is overrated for most product use cases.
Better retrieval, better tool design, and tighter feedback loops usually move the needle more than another round of training

[Prabhat Singh @prabhatxz]

Context windows keep getting longer, but most systems still treat memory like a dumb buffer.
The real unlock is selective, structured memory that the model can actually reason over across sessions

[Prabhat Singh @prabhatxz]

@amritwt Indeed it is a hassle, a good alternative I am using is infisical

[Prabhat Singh @prabhatxz]

@amritwt you are on discord?

[Prabhat Singh @prabhatxz]

@archiemckenzie_ why do you think so?

[Prabhat Singh @prabhatxz]

@jatingargiitk it is true, but how will you know before building whether people wants your product or not?

[Prabhat Singh @prabhatxz]

@dreidel1 how will your country make most out of it?

[Prabhat Singh @prabhatxz]

@rauchg ai is going to take over then, so what will you do when you will be free from all the tech work?

[Prabhat Singh @prabhatxz]

@Amitesh48256 neovim has always been very good in terms of using the least of the resources unlike VS code which is built upon Electron which just uses much amount of RAM

[Prabhat Singh @prabhatxz]

Most agent frameworks focus on planning and tool calling.
The missing piece for real products is still evaluation and observability — knowing when the agent is quietly failing before users complain

[Prabhat Singh @prabhatxz]

The biggest gap in current agents isn’t intelligence
It’s reliable long-horizon execution
They can plan for 5 steps
They still collapse when the 6th step depends on imperfect earlier output