Srinathprasanna

Fintech Enthusiast | Passionate Blogger | Avid Reader | Lover of Mother Nature and Earth

[Srinathprasanna @Srinathpras]

Open source.. runs on your box, the text never leaves it.

Repo + the reproducible video check:

https://t.co/Lx7RKtrvBX

If you can show me where the recount is wrong, that's the feedback I want most. @GawkDev

[Srinathprasanna @Srinathpras]

Your 5 second AI video cost 246,840 tokens. I rederived that exact number from the output file alone.. no trust in the invoice. Delta: 0.

Here's how token bills are (and aren't) verifiable 🧵

[Srinathprasanna @Srinathpras]

What it does NOT claim: that anyone is overbilling (my counts matched), or that the price is fair. It checks the count. The point is you can finally check it yourself.

vs @LiteLLM / @helicone_ai : they aggregate the provider's number. This re-counts it.

[Srinathprasanna @Srinathpras]

@bookwormengr completely agree to this @bookwormengr . I completely believe open weight models are the way forward. The amount of marketing push by the western models make us rethink anything of Chinese origin. we should not forget that America outsources most of their chipfab works to china.

[GDP @bookwormengr]

India needs to put ego away and embrace global open source AI
================
Indian policy makers and wider industry appears to have issues in using models trained in China. That is like throwing baby out with bathwater.

China based labs have been source of world' best Open Source AI. 81% on Terminal bench by GLM-5.2 is simply mind blowning. It is a very hard to master benchmark. What is more is this Anthropic Opus level model is Open source. Open source AI is not stopping there and will keep improving..There are great many western options also emerging but Chinese labs stay class apart.

What about these models having Chinese world view? Ain't China a competitor? Ans: You can always post train models and put your world view into them. Also, when you are using model for coding, scientific and engineering works, you don't have to ask the model "what happened in Tianmen square". Western models also have their own biases.

What about data safety? Ans: When you run models under your control, you don't have to worry about data being sent outside the country.

You don't have to wait for Sarvam to build all models. I love Sarvam and what they are doing - but they don't build model for every use case and category (at least for now). There are high quality open source models are available for different scenarios. No need to wait more.

C-DAC for Inferencing and post training:
================
Every journey of thousand miles stars with the first step.

You don't need to master pre-training as the first step. It is hard. Indian industries can master the art of running high class inference themselves (follow that with post training and RL). It is very different from naively deploying models. There are great collaborators like @radixark @inferact @NVIDIAAIInfra who can teach you how. It is a challenging worthy of your effort....

India government needs to do it for highly privacy sensitive use cases (e.g. defense), but beyond that for other use cases too Indian government could have a C-DAC like agency to manage inferencing and post training. Use OpenAI, use Anthropic where you must; but also maintain your own inference & post training muscle. Open source AI is so good, you could practically use it for 95% use cases.

Also, you need to master how to setup data centers and clusters. There is a lot to be learnt.

For private industry question worth pondering: where lies the value?
================
When models become commodity - data and post training gain all the value. Own the data, and post training stack. Protect your IP. Please read this brilliant essay by India's very own son Satya Nadella https://t.co/CzzqsNrwsg

Furthermore, self hosting allows you to collect data as users use the system and continuously improve your models. This is called Online RL. For example, Cursor - Coding startup that SpaceX acquired for 60B - started with Kimi models from Moonshot and improved it with post training (RL). Then they applied Online RL, updating models every 5 hours. They collected data on how their users were using the model and continuously updated the model. The results are for everyone to marvel at. They were helped by @FireworksAI_HQ . Lot of help is available if you look around. You can also work with Hyperscalars to master the art of high scale inference and post training - e.g. AWS's Hyperpod offering.

You just need to have the spirit of late JRD Tata of leading from the front. The defeatist attitude shared by some IT industry veterans is totally unwarranted.

Also, in future you are not assured to have access to top class closed source models due to various security and competitive consideration. Start planning for that future when your global competition may have access to models that you don't have access to.

No room for self doubt:
==============
- China's mobile maker Xiomi makes great cars, and now builds world class LLMs. Inferencing and post training is much easier than that.
- China's food delivery company Meituan too has amazing skillset in build world class models.

No room for self doubt : model weights, code, infrastructure know how is available in open domain. You can do 95%+ work with open source AI models. If you do it properly you can also save massive $$$.

It requires strong will - I hope you all have it. Remember what JRD did for India, you could do it for AI.

@svembu @Krutrim @bhash

[Srinathprasanna @Srinathpras]

The product is best in class. The support isn't. Fix the bot or route paying customers to humans. Simple.

[Srinathprasanna @Srinathpras]

Paying £90/month for @claudeai Max. Reported a service outage with screenshots. Asked two simple questions.
Got an AI generated reply rejecting a refund I never requested. Clarified the mistake. Conversation auto closed as "resolved."

#Claude #Anthropic #AISupport #GenAI

[Srinathprasanna @Srinathpras]

An AI company whose AI support can't read a two paragraph email.. not a great look.

[Movez @0xMovez]

Jane Street AI Engineer revealed how they trained their own LLM for trading to make $22.5B/year

16 minutes. free. straight from tier-1 quants.

bookmark & watch - this is the most honest "AI inside a hedge fund" talk ever published.

forget the "AI trading bot" YouTube grifters. This is the real inside view: data, training, evals, integration.

then start building your own bot using post below.

[Srinathprasanna @Srinathpras]

News: MTP is all about acceptance rate — via Gawk https://t.co/AKNBU9ndRN

[Srinathprasanna @Srinathpras]

Your AI tools have a trust problem. You pay for them. You build on them. You ship production code with them. But when something breaks, you check Twitter or X to find out if it's you or them.

That's not engineering. That's guessing.

https://t.co/RjPltnUATE gives you one feed.

[Sridhar Vembu @svembu]

I studied the life and work of Lee Kuan Yew, the Founder-Prime Minister of Singapore intensely during my early 20s. Immense respect for him.

In this video clip, he advises a young woman pursuing a PhD not to neglect marriage and kids. I recently gave very similar advice.

Elon Musk would say much the same thing.

We have to orient our culture, society and the economy so that smart ambitious women do want to have children (and sadly, this needs to be said today: men cannot have children). If we fail at this, humanity will lose.

Now let the attacks begin! 🤓

[Srinathprasanna @Srinathpras]

@P9Ashwini True and resonates well with every situation in life

[Srinathprasanna @Srinathpras]

@RCBTweets Check out .. https://t.co/OW9SWKQnWw as well

[Srinathprasanna @Srinathpras]

124 Repositories. Four Ecosystems. One Broken Assumption.

https://t.co/JPmiWiHbMw

#Gmail
#NextAuth
#PassportJS
#SelfHosted
#GitHub

[Srinathprasanna @Srinathpras]

Google let people change their Gmail address. Their own dev docs admit it creates duplicate accounts and data loss in downstream OAuth integrations.

Nobody's talking about it.

So I wrote the technical follow through, ran a code audit across 2M+ repos and built a free scanner.

[Srinathprasanna @Srinathpras]

The tool: pip install authdrift && authdrift scan ./

Finds every OAuth handler in your codebase that keys on email instead of sub. MIT licensed, runs in CI.

https://t.co/EfEntWgIS6

#infosec #oauth #google #gmail #appsec #identitysecurity