Aakash Kambuj

Something new. Previously - AI architect at Highspot, founder/cto at Falkon AI (sold to Highspot), engineering at Dropbox, Familiar and Microsoft.

[Aakash Kambuj @Aakashk]

A natural fit for this seems to be a Pocket/ReadItLater class tool. But more AI native. Have you run into this issue? Found anything that works?

[Aakash Kambuj @Aakashk]

I frequently run into this issue. An LLM conversation will flag something interesting, and I want to save/persist that message (and possibly some surrounding context) for the long run - not only for the AI but as reference material for me in the future.

[Aakash Kambuj @Aakashk]

This is pretty cool.

[The Bugged Dev @thebuggeddev]

Only if education could be this interactive ❤️‍🔥

I've had a looong wish to build something genuinely useful through vibe coding, and I finally did it.

A 3D human anatomy application built with @threejs using GPT 5.6 Sol.

It all started with a single design image that I created using GPT Image 2.0. I then used it to generate every 3D organ image, one by one.

Next, I converted each of those images into 3D models using @tripoai (and no, they didn't sponsor this 😄).

After that, I opened Codex, wrote a master prompt based on the design, and gave it the prompt, the design image, and all the 3D models.

Codex built the first version beautifully, but there was one big problem.

Every single 3D model was nearly 120-150 MB. That obviously wasn't practical for the web and was giving a performance of 16fps.

After a few iterations, Codex optimized each model down to roughly 2–5.5 MB while preserving the visual quality, reducing the total asset size from ~900 MB to just 28.6 MB. And each model loads on demand.

Along the way, Codex also generated those anatomical illustrations showing where each organ sits in the human body, and even created the interactive hotspot markers that explain different parts of every organ. It handled all of that.

The process wasn't exactly one shot, but it also wasn't difficult. You just have to do it step by step.

It genuinely felt like building something that could make learning anatomy much more engaging.

The inspiration came from @DilumSanjaya's 3D animal plant cell project. I remember seeing it and thinking, "I want to build something like this one day."

And I did it :D

Live: https://t.co/gIFgkj8UVB
Code: https://t.co/nijvyWcyxO

[Aakash Kambuj @Aakashk]

Alternatively- frontier labs hold back true frontier capability from the API entirely, serving it only through products where outputs can’t be harvested at scale. Agentic product offerings.

[Aakash Kambuj @Aakashk]

Inference dominating training implies a razor and blades model — open weight model is then razor (one time training cost), selling inference against it is the blades (leveraging the deployed silicon and electricity).

[Aakash Kambuj @Aakashk]

US cos can fine-tune enterprise models on the N-1 version, which would solve for any kind of regulatory need to keep using US models, while getting the price benefits of an OSS model (assuming, of course, the N-1 model is price-competitive to serve and not a fuel-guzzler V12)

[Aakash Kambuj @Aakashk]

Frontier labs open-weighting in USA feels so unlikely though. Maybe we get a rolling release valve, where each cohort of weights opens once its successor ships and its distillation value has decayed, keeping the ecosystem seeded while the premium wedge stays exclusive.

[Aakash Kambuj @Aakashk]

3. Wildcard - a government-bankroll scenario - where a subsidized lab no longer needs weight exclusivity for revenue and Washington decides seeding the world with American weights beats ceding the Global South stack to Qwen or Kimi.

[Aakash Kambuj @Aakashk]

2. the quality gap to Chinese OSS closes to the point where the closed premium can’t cover serving costs, at which point weights have no exchange value anyway and opening them is a free ecosystem play. Wonder if we are close to this.

[Aakash Kambuj @Aakashk]

US labs open-weighting frontier models could happen under specific triggers:
1. a lab falls decisively behind and flips to the spoiler strategy.

[Aakash Kambuj @Aakashk]

Open-weighting frontier models is a weapon of the second-place player against the first — which is exactly what it is for China nationally. Meta’s entire Llama strategy was this: no serving-exclusivity to protect, so open the weights & commoditize the leader’s advantage.

[Aakash Kambuj @Aakashk]

The lab becomes a studio; the cloud becomes the theater chain; and the theater chain always ends up owning the studio, because it’s the one with the balance sheet and the distribution.

[Aakash Kambuj @Aakashk]

Training is customer-acquisition cost for inference revenue

[Aakash Kambuj @Aakashk]

This should devolve into everyone “giving” away their models as open weights, and hyperscalers + neoclouds + specialized serving teams (fireworks/baseten/together ai) win revenue (they possibly bankroll future training runs for their favorite labs)

[Aakash Kambuj @Aakashk]

@MelvinInvests Contrarian take — MU pricing is predatory and unsustainable for the health of the broader AI ecosystem. And could strangle its own revenue growth in the future.

[Aakash Kambuj @Aakashk]

The CAP theorem of home construction - at most you will get two out of these three variables:
- On time
- On budget
- With high quality

I call this the Hard Hat theorem 🧑‍🏭

[Frank Downing @downingARK]

https://t.co/najMqP3mxN

[Aakash Kambuj @Aakashk]

Both companies and humans want to preserve their unique IP

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Aakash Kambuj @Aakashk]

The $SPCX approved 2026 version of "to the moon" is "to the stars"

[Aakash Kambuj @Aakashk]

But hey I'm not complaining, it's probably gonna pay for my Claude Max for the month...

[Aakash Kambuj @Aakashk]

Anyone remember AllAdvantage from the dotcom boom days? "Get paid to surf the web" ?

Just came across https://t.co/2tC4PqvKRF and I've already made $8.50 in an hour. Certainly feels like peak AI now...