Building AI that closes deals for SMEs Founder · Operator · Builder
Joined October 2009
- Tweets1K
- Following2.1K
- Followers182
- Likes4.4K
LLMs predict the next word.
World models predict the next consequence.
Text tells you the glass breaks. A world model has to predict where the pieces land.
That's why robotaxis, robots, and soon your AI agents need one.
ALT LLMs predict the next word. World models predict the next consequence.
"Your price is too high."
The 3-rep drill:
1. Clarify, don't defend: "Too high compared to what?"
2. Anchor to their cost of the problem, not your price
3. Trade, don't discount
3 reps x 3 buyer personas. 15 minutes. Weekly. #SalesIsASport
ALT Your price is too high: the 3-rep drill.
Spent 9 hours building Claude Code skills instead of hiring a motion-graphics agency.
Now every FireCoach launch video is: script in, on-brand animation out, tweak in plain English, render.
The skill is the asset. The video is just output.
ALT 9 hours of Claude Code skills replaced a motion-graphics agency.
Rivian just bought Third Dimension, a 15-person world-model startup.
Not for games. For robotaxis.
Its fleet uploads 400k+ driving sequences a day. A world model makes that data simulatable.
LLMs predict the next word. World models predict the next state of the world.
ALT Rivian just bought a world model.
Recording sales calls isn't coaching. It's game tape.
NBA teams watch tape AND run drills. Most sales teams only watch.
Reps get better by running the hard conversation 10x before a real buyer hears it.
That's why we built FireCoach. #SalesIsASport
ALT Game tape isn't practice.
Not every automation needs an agent.
Known steps, clean input: plain workflow
Known steps, messy input: workflow + one LLM step
Unknown path: agent
Can you draw the flowchart? Then you don't need an agent.
ALT Not every automation needs an agent.
I'm 8 drills from being certified as an SDR. By my own product.
Avg score 69. Pass mark 65. Day streak 2.
Next up: Skeptical Sam, Budget-Blocked Brian, Status Quo Steve.
Drills before games. That's how a ramp should work.
Ramping SDRs? Reply RAMP.
ALT FireCoach ramp: 8 drills from certified.
The SDR isn't dying. The SDR team is.
Old: 5 SDRs blasting 400 emails a day.
New: 1 GTM engineer builds the system. 2 SDRs talk to the 40 accounts actually in-market.
Fewer sends. More meetings.
But automate a bad ICP and you just spam faster.
ALT Your next SDR hire might not be an SDR.
Built an AI sales coach. Then let it coach me.
Call vs Gatekeeper Grace: 69% readiness.
Win: quantified the value at 0:59.
Miss: 13% talk ratio (target 30-55%), and never asked who else decides.
Running it back tomorrow.
Ramping SDRs? Reply RAMP.
ALT FireCoach call report: Gatekeeper Grace, 69% readiness.
Harry Kane just won the Ballon d'Or.
Years of finishing 10th, then 13th. Then 73 goals in a season.
Not luck. Reps.
Same in sales. Top closers are built through practice, not born. That's what FireCoach does.
Who's the Harry Kane on your team?
#FireCoach #BallonDor
World models series, done. Start here:
1. LeCun, A Path Towards Autonomous Machine Intelligence (2022)
2. I-JEPA (2023)
3. V-JEPA 2 (2025)
Want to build, not just read? IIDA's JEPA track needs a lead + contributors. DM me. 10/10
ALT World models reading list and call for a JEPA track lead
How to get involved with IIDA:
Low lift: intros, compute credits, pro-bono help
Medium: 2–3 hrs/wk as pod lead or co-author
All in: Research Lead, Program Manager, patron
Or bring a problem worth solving.
Innovation out of India, on the record.
ALT Low lift to all in: ways to get involved with IIDA
AI agents aren't blocked by models. They're blocked by interfaces people can trust, correct and approve.
IIDA's Agentic UX track works on that layer, the adoption bottleneck nobody owns.
Builders: what's your hardest agent UX problem?
ALT Black-box agent vs legible agent with steps and an approve button
IIDA Track 02: World Models and JEPA.
Seeking a lead. 6-10 contributors. Written kill criteria.
The question: which Indian SME problems could a JEPA-style model solve in 3-5 years, and with what data?
9/10
ALT IIDA Track 02: World Models and JEPA, seeking a lead
Our name says All Bharat. Our map says Pune.
ABAIA plots every member by city on its homepage. Almost every dot sits on one point.
The gap between the name and the map is the work.
Building AI outside Pune? Take a seat.
ALT Member dots by city: mostly Pune, empty seats elsewhere in India
IIDA target: ≥30% of revenue from industry contracts by Year 2.
Fraunhofer's discipline, not a grant habit.
Got an applied AI problem your team can't staff? The problem bank is open.
ALT 30% of revenue from industry contracts by Year 2
Two roads to world models:
Generative: predict the next frame/token
JEPA: predict representations, skip the unpredictable
No consensus yet. Open questions are where new labs can contribute.
Which would you bet on? 8/10
ALT Two roads to world models: generative vs JEPA
IIDA's top 3 risks, published:
1. Credibility cold start → ship a visible artefact in 90 days
2. Funding before proof → raise compute first
3. Scope creep → kill criteria + one focus by Q4
Five more on the site.
ALT IIDA risk table: risk, why it's real, our move
Reply REPS. I'll build your team a practice call for your toughest objection.
Your team should practice before the call that matters. Not on a real lead.
Before the shift, 10 minutes, against an AI buyer that interrupts and haggles on price. That's FireCoach.
ALT Practice before the call that matters