PC

Chat was never where work lived. Building @withlemma → the place AI agents and humans actually work as one team. open-source. Founder, building in public.

[PC @hey_its_pc]

@kapeed_jha @withlemma I lost myself Between "how can something so freakishly good be opensource" and "wait! This not a @ycombinator product"

[Deepak @kapeed_jha]

This was the first time we took @withlemma out to the world. And boy, did people love it...

[PC @hey_its_pc]

@iamrealsnow mine shit the bed the moment a human has to do physical movements and not camera pans

[PC @hey_its_pc]

only if influencers who have been writing video editor obituaries actually created stuff they promote
there goes 35$ to @runwayml - 7 attempts - each more absurd than the previous one

[PC @hey_its_pc]

agents are already doing employee-level work. Why are you still treating them like a visitor? Try lemma - its open source - has all the open source LLMs to keep you running and also works with your claude and chatgpt subscription!
@withlemma

[PC @hey_its_pc]

@Noahpinion Canada made the rookie mistake of trying to grow GDP without periodically putting a country on the regime-change roadmap.

[Ayush Gautam @_ayush_gautam_]

LFG 🚀

[Deepak @kapeed_jha]

Eventually everyone AI product will do this in some way or another.
- Spin up some tables for you track things you might want to query and track long term
- Files for general playbooks (skills) and anything that doesn't fit in structured data
- Agents as the execution
- iMessages, Whatsapp, Telegram, Slack, Custom Apps as clients

[PC @hey_its_pc]

@SurKopu The part I'd add: for small teams the new bottleneck is review bandwidth. Agents can produce ten times the work, but a human still has to approve the parts that touch money, customers or prod. Taste shows up as knowing which steps need a human and which don't.

[PC @hey_its_pc]

@BMadusudanan Agree, though I'd narrow it: the coding experience pays off at review time, not prompt time. Knowing which step to gate and what a wrong intermediate result looks like is what keeps an agent loop from quietly doing damage.

[PC @hey_its_pc]

👏🏻 Oh yes! We open sourced @withlemma recently, and one of the recent grads building with it sent over his first Dodo Payments receipt a couple days back.

For a second, the grind suddenly felt worth it. ❤️

Tldr of his build: He built an app + WhatsApp agents that collect expenses over WhatsApp, flag discrepancies, add everything to a dashboard, and automatically follow up based on the org’s escalation rules.

[PC @hey_its_pc]

@deepseek_ai Flash V4 is now doing real work inside some of our @withLemma apps. output holds up. cost barely registers.

That changed the question for us.
It’s no longer: “Which model is smartest?”
Instead: “Did we design the work properly?”

Small models for routine work. Strong models for judgment. Lemma gives both the same state, permissions, workflows, approvals, and team interface.
The model matters less when the architecture is doing its job.

Lemma is open source. Repo in the first comment.

[Deepak @kapeed_jha]

Model is becoming the least interesting part of an AI product. We are running real work inside our @withLemma apps on DeepSeek Flash V4. The results are good. The cost is tiny.

Because most business work does not require a frontier model.

It requires good state, clear permissions, deterministic transitions, human approvals, and an interface the team can share.

The winning architecture will not run every task through the smartest model.

It will know exactly when intelligence is needed - and when it isn’t.

@withLemma makes this possible - not sending every task to the smartest model, but letting me decide which agents are

Lemma is open source. Repo in the first comment for anyone building internal tools or personal productivity apps.

[PC @hey_its_pc]

And it runs on what you already pay for.
Any model per agent, your keys, your machine.
That's @withlemma .
Open source.Describe how your team works. One command makes it real.
https://t.co/SC6KMgK7hi

[PC @hey_its_pc]

The software my team actually needs didn't exist.
We open-sourced @withlemma for exactly this - so teams can have personalised workspaces for humans + Agents - and work stopped getting lost in chat.
👇🏻Here's one I use the most and
🧵How I built it in <2 hours

What it is: agents + people on one roster.
Agents own tasks. Nothing ships without a human signing off.

[PC @hey_its_pc]

Same agent, wherever my team already is.
assign from WhatsApp → it lands on the board.
Slack's the same agent again.
All surfaces are fully loaded.

[PC @hey_its_pc]

🌊 clever comparison, also a deeply unsettling one

👆arithmetic = correct
But the accounting, they would give a cabinet seat in Indian govt for this - not “apples to apples”

🍔 hamburger footprint = (full supply-chain water) rainfall absorbed by crops+ irrigation water + theoretical quantity of water required to dilute pollution

AI-image footprint = water for operating data centres & generating electricity.

What we miss: footprint changes dramatically depending on the model, grid, weather, location and time of day.

A litre of rain falling on cattle feed is not interchangeable with a litre of freshwater consumed by a data centre in a water-stressed district.

Water is not just a quantity. It has a place, time, owner and an opportunity cost.

“golf courses consume more” is not a defence but environmental equivalent of a child saying: “Why are you questioning me? He used more”

existence of a larger problem does not magically dissolve a smaller one
AI infra is expanding rapidly - its costs can be concentrated in specific communities.

I have spent the last three years building with AI. I believe deeply in what this technology can do. That is precisely why comparisons like this bother me.

We should be asking:

💧 What kind of water is being consumed?
📍 Where is it being consumed?
☀️ Is that region already water-stressed?
🏭 Who benefits from the data centre?
🏠 Who absorbs its environmental cost?
📊 Are companies even disclosing enough for us to know?

This becomes even more uncomfortable coming from an Indian founder.

As we debate whether a hamburger equals 83,000 AI images, Adivasi women affected by the Ken–Betwa project have been lying on symbolic funeral pyres with protesters also using nooses to demand justice over land, displacement, compensation and rehabilitation. Development may appear as national aggregate on one spreadsheet. On the ground, it lands on somebody’s home.

Here is the philosophical problem with this entire style of argument:

🌳 Suppose planting a tree creates as much measurable “good” as one person creates.
Can I now kill a person and plant a tree?
Obviously not.