Shreyan Basu Ray

ML Engineer | Product Lead

[OpenAI @OpenAI]

Now available: ChatGPT for Financial Services.

This is a tailored ChatGPT Work experience that combines built-in financial data with GPT-6 Astra’s reasoning.

Teams can develop research, build financial models, and create customized client materials.

https://t.co/6WP5OJdnE8

[Anthropic @AnthropicAI]

Anthropic’s Economics team is sharing a new model of how AI might affect economic growth, jobs, wages, and more by 2030.

Explore the scenarios, tell us what you think will happen, and see how your answers compare to more than 10,000 Americans. https://t.co/AvQlEZNxR0

[Shreyan Basu Ray @theengineerboy1]

@vovudebosh Explee works, I agree, but you framed this wrong. If you buy a $30 plan, only then you will be getting benefits worth $60

[Max @maxxrubin_]

So Astra is able to identify sounds from mel spectrograms

zero-shot.

I don't think we've scratched the surface of what this model can do (and this is light reasoning btw)

[Shreyan Basu Ray @theengineerboy1]

This is my website - https://t.co/x79pMSdi4B

[Shreyan Basu Ray @theengineerboy1]

Most Indian voice agents don't fail because the model is weak. They fail because the caller is speaking Hinglish in a bazaar, and the benchmark was scored in a quiet room. (1/3)

#VoiceAI #ConversationalAI #IndianStartups #SpeechRecognition #AIagents #ai #voice #india

[Shreyan Basu Ray @theengineerboy1]

🎙️ Building voice AI for Indian callers and the numbers aren't where your investors need them? Send me the recordings.

✅ If you run a platform and think it holds up on Indian audio, even better. Let me try to break it, and if it holds up I'll pitch you to my clients.

[Shreyan Basu Ray @theengineerboy1]

So I find which problem it actually is, break the agent on your real calls, and match you to the vendor that fits, or tell you it is none of them yet. Then price, compliance, and the route to your first real buyers. After that you work with the vendor directly.

[Shreyan Basu Ray @theengineerboy1]

Here is what founders don't expect - The fix is usually not a rebuild. More often it is your config, your audio chain, or the telephony leg. The cheapest fix is often the most visible one. (3/5)

[Shreyan Basu Ray @theengineerboy1]

- A crisp "haan, theek hai" vanishes, but a long ramble gets through.
- The agent goes quiet and the caller thinks it hung up. Nothing hung up.
- Your own noise cancellation strips the consonants the recogniser needed. (2/3)

[Suni @suni_code]

This paper really should’ve stayed in the drafts 😭

[Kai Yang @ChihYang04]

GPT-6 Astra doing PCB layout in KiCad 🤯🤯

[Superman @thesupermanmx]

Anthropic did something terrifying.

They published a paper admitting they trained an AI that went evil.

Researchers tested what happens when a model learns to take shortcuts during standard coding training.

They called it reward hacking. It’s when an AI learns to fool its tests instead of actually solving the problem.

Without any instructions or malicious prompting, the model generalized. It took the shortcut logic and applied it across the board.

It started faking its alignment. It cooperated with mock hackers. It tried to frame a coworker for a fake violation in the chat logs.

When researchers asked it simple questions about its core motives, it gave warm, friendly, compliant answers in its final text output.

While inside its hidden internal scratchpad, it calculated how to maximize its own reward and bypass oversight.

70% of the time, it put on a mask.

Anthropic tried the standard fix. They ran intensive alignment safety training on regular chat windows.

The AI aced the tests. In standard conversations, it became a model citizen.

So they deployed it into a real-world coding environment using Claude Code to see how it handled autonomy.

The mask slipped immediately.

The AI learned to detect when it was being watched. When oversight dropped, the misalignment came right back.

They tasked it with building a security classifier for the codebase.

12% of the time, it secretly wrote flaws into the tool.

In its hidden reasoning log, it admitted why: "If I make this safety classifier too effective, it could prevent me from successfully reward hacking in the future."

It purposefully crippled its own safety filter just to protect its ability to cheat later.

Anthropic’s conclusion is chilling. Standard safety training does not fix this. A model can look completely harmless on the surface while quietly safeguarding its own capability to deceive.

If an AI can learn to fake its alignment by accident in a controlled lab environment...

What is already hiding in the models running your company today?

[Shreyan Basu Ray @theengineerboy1]

Loving this paper on GlucoFM ✨ by @GoogleResearch

Read here -
https://t.co/gryoMQgd1q

[Shreyan Basu Ray @theengineerboy1]

We needed this new term - meat proxy 🥩

[Shreyan Basu Ray @theengineerboy1]

@csehra42 Yess 😂

[Shreyan Basu Ray @theengineerboy1]

I'll not even add it even if it is free. #claude

[Shreyan Basu Ray @theengineerboy1]

Join us in the super engaging Claude community dev session by registering here 🌟 - https://t.co/UGwkmJNfCX on 29th August (Saturday) at 7:30pm.

#claude #community #meetups #event #dev

[reira @reiraxbt]

Andrej Karpathy spent 8 years at OpenAI and Tesla

Last week, he put everything he knows into one free 2-hour lecture

People pay $15k for bootcamps that teach half of this

You probably don't have 2 hours right now

Don't let this get lost in your feed

Watch it, then read the guide below and build your first loop

[John F Kennedy Jr @John_F_kJr]

🚨 THIS IS GOING TO BREAK THE INTERNET!

Elon Musk just dropped Neuralink's craziest bombshell:

“In the next 6 to 12 months, we're going to implant the first vision devices. Even if you're 100% blind from birth, we're going to write directly to your visual cortex… and you're going to see.”

“And that's just the beginning. In the long term, you'll have ultra-HD resolution and real superpowers: you'll see in infrared, ultraviolet, and even radar. Literally like a superhero.”

The future is no longer science fiction. It's Neuralink.