Srikanth Varma Chekuri (@cskvarma) header
Srikanth Varma Chekuri (@cskvarma) profile picture

Entrepreneur (1 Exit, 1 Failed, 2 WIP). Applied AI Scientist (18 yrs at Amazon, Yahoo, Startups, IISc). Loves teaching Math and AI.

Hyderabad, India
Joined October 2009
Any company that values it's data, internal know-how and IP would stop using Anthropic or OpenAI for critical work, and simply self host open-models. It makes no sense when these companies can come for you, next quarter or year.
3
3
40
1,547
Here is an interesting challenge to tackle, assuming Software, Math and Engineering is solved by SOTA AI: 1. Make a sub 10B model work as well as SOTA AI for knowledge work, on consumer laptops. 2. If (1) fails, prove the foundational Math on why it cant be done? ........(1/2)
4
27
3,244
Solving (1) makes it universally accessible and kills the AI hype. Solving (2) gives us a fundamental understanding of intelligence. .......(2/2)
8
429
I thought deeply on what drives me to build, read, write, learn, mentor and work….I think I have a reasonable answer, for now…”the joy of figuring things out”…. I feel blessed.
1
40
1,209
Lets be honest with ourselves: the way we do engineering or science or any intellectual work has changed forever, whether we like it or not. Leverage it or be left behind. I feel liberated, as a lot of things that were impossible for me a few years ago, suddenly became possible.
7
100
2,375
Every engineer or consultant at a services firm has to become a Forward Deployed Engineer (FDE). Sit with the customer and build what they need. Make your customer 10x efficient.
5
4
70
2,444
Maximize value-producing and minimize unnecessary consuming….
1
19
1,095
Should sovereign AI models be built from scratch? Is it sensible to design a fighter jet engine from scratch if we get working open source designs? Shouldn’t we just modify the models/designs for our needs, to speed up.
1
15
1,385
GenAI is closest to finding a “cheat-code” in games, just that the game is now real.
7
466
If you have something that others want/need and you can profitably trade with them, you become a prosperous person/region/country, as long as you can protect it. It's that simple, whether it is ancient tin, iron, spices, silk, ships, guns, opium, skilled workforce or siliconchips
1
20
763
College faculty must spend 25-30% of thier time building real and hard things. Consulatants must spend 25-30% of thier time generating real outcomes. Founders must spend 25-30% of thier time doing every job in the team, to really know the nuts and bolts.
2
1
35
1,320
Make kids design hard math problems which they think are "hard". Just like teaching is the great way to learn, designing hard problems is a great way for them to be good problem solvers --- lessons from the trenches of working with a 8 yr old.
18
670
If batteries explode, the onus lies with the battery maker. If a cloud DB has a bug which leaks data to hackers, the cloud provider is liable. What makes closed source AI mishaps any different? With open source, the onus lies with the InfoSec team of the enterprise deploying it.
2
24
724
The best option for OpenAI, Anthropic, Gemini and Grok seems to be: 1. Open source and lower the inference cost of Tier 2 and 3 models to build enterprise trust and cater to price sensitivity. 2. Tier 1 stays premium and gets updated every 3-6 months as open-weights catches up.
1
1
19
1,014
Few things that people miss in AI assited learning (any topic): 1. Struggle creates skills. You need to struggle while solving problems to build comptence. There is no shortcut. 2. AI to be used as a great-teacher which unlocks minimally when stuck, and not remove the struggle.
3
3
68
3,044
In the age of AI, the question to ask oneself is: what is the hardest problem (that I care about) I can solve today, using AI. Sit on it. The day you solve it, rest and level up to the next one.
1
39
1,205
Cloud succeeded because AWS did not peep into the internals of Netflix, Slack or Airbnb. “Trust” was earned and kept intact. Same cannot be said about OpenAI and Anthropic or any closed weight AI model provider. (1/2)
34
469
23
7,042
253,690
I foresee a flood of enterprises, researchers and government organizations and institutions moving to cloud or local self hosted open models. The same will happen for every economically and intellectually useful application of AI: from Math to Drug discovery.(2/2)
6
43
1
1,293
27,100
OpenAI and Anthropic lack “trust”, which is what matters in the medium to long term. Navier stokes story, hugging face hack, Anthropic’s push for AI regulation are some examples. I much prefer open models that can run on hardware I chose, with privacy settings I want.
2
25
1,119