Debjyoti

Developer | Socially awkward and I use arch btw

[Debjyoti @Debjyoti0404]

What if the real AI breakthrough isn’t scaling compute, but decoding how the human brain thinks?

​LLMs burn megawatts to reason, while the brain runs on ~20 watts.

​Are we brute-forcing the wrong architecture, or will scaling eventually bridge that gap?

[Debjyoti @Debjyoti0404]

Might be late, but exploring Clairic a System 1 extraction model
Unlike JEV or Clef(fixed choices), Clairic extracts typed values & forecasts distributions in <1s.

Key returns: span offsets (provenance), p_null (missing data), & p10/p90

Full blog: https://t.co/ZjUeNWVr6r

[Debjyoti @Debjyoti0404]

Relatable enough😶

[Debjyoti @Debjyoti0404]

I mostly use subagents as the generators with each agent handed bite sized works, then the supervisor evaluates the implementation as it has enough context to check for edge cases and the loop continues.

[Addy Osmani @addyosmani]

Before adding another subagent ask what problem it's solving. A bloated system prompt is often the issue.

Skills and tool search can fix that without splitting context across agents. My rule of thumb is to keep a subagent when the work is parallel, needs fresh context or one agent really can't pick the right tools.

Otherwise, fold it into the main loop, turn its prompt into a skill, run your evals and keep whatever wins.

Good write-up:
https://t.co/1gcrNNpQru

[Debjyoti @Debjyoti0404]

AI is already capable enough to quickly build microsaas based on niche requirement by the people who need it, then what is the point of building SaaS in the current time?

[Debjyoti @Debjyoti0404]

@Anjishnu46 @cneuralnetwork I think the current state the models are in, can already write and decide better design principles then many of us.

[Debjyoti @Debjyoti0404]

As much I am bullish on agentic workflows and scope of new automations, news like this really shakes up the confidence on the industry we are currently in, maybe it is coming for everyone eventually

[The Kobeissi Letter @KobeissiLetter]

BREAKING: Microsoft, $MSFT, is offering buyouts to 60,000 workers as part of the company’s efforts to cut costs amid AI adoption, per Bloomberg.

Personnel on select teams in the US, Canada, India, and the UK at a vice president level and below will be eligible.

This reflects ~27% of Microsoft's total workforce.

[Debjyoti @Debjyoti0404]

@piyushgarg_dev But to get that 1 project you need to build 1000

[Debjyoti @Debjyoti0404]

@cneuralnetwork I haven't used other harness beside claude code, what is your opinion on DeepSeek model? Planning to get started with it in opencode

[Debjyoti @Debjyoti0404]

@cloneisjun Yes definitely, what I am planning is to create some unit tests based on the gathered proof by the supervisor agent and the subagents will be tasked to pass those tests. Otherwise vague task description does not give us verifiable results if we want to scale this setup

[Debjyoti @Debjyoti0404]

I was planning to implement a auto bug fix workflow at my work where a agent understands the defect, gathers proofs, produces it and hands off tasks to certain subagents.

Stumbled upon this blog, it was a good read indeed @dexhorthy
https://t.co/fHnyJXCivH

[Debjyoti @Debjyoti0404]

@ThePeterMick In the early 2024 we said we are not going to trust AI to own full access of our codebases and used them as auto completion and suggestion. But here we are now. Tech really moves freakin fast

[Debjyoti @Debjyoti0404]

@loftwah Maybe the problem solving mind we develop by learning code will be helpful as problem will be there, only the tools to solve the problems have just changed

[Debjyoti @Debjyoti0404]

@mehulmpt Nowadays reading git commit messages are nightmare itself, some days ago my co worker pushed 2 paragraphs as commit message😭

[Debjyoti @Debjyoti0404]

@nandanchillal Madara was right about life

[Debjyoti @Debjyoti0404]

@md_kasif_uddin The shift from treating LLM on 2023 like god on answering basic general knowledge questions to cursing it on why it couldn't one shot sqlite engine in Rust we came a long way 😐

[Debjyoti @Debjyoti0404]

@hayyantechtalks Crazy times

[Debjyoti @Debjyoti0404]

@arpit_bhayani Is true software factory possible where the agents itself decides the architecture, code, tests based on new product requirement or human will still be needed to maintain the system in the long term?

[Debjyoti @Debjyoti0404]

@JoschuaBuilds I am also getting this suggestion often. Don't know how beneficial it is in terms of growth but the strategy is reducing my social anxiety of interacting with someone.

[Debjyoti @Debjyoti0404]

Tldr version

[Debjyoti @Debjyoti0404]

@sankhadiproy23 So far what I have understood is in case of a traditional serverless model like lambda a new container is started during a cold start which takes a significant amount of time and lambda can only process a single request at a time.