Bishal Guha Mallick

Leading AI Initiatives @ Internshala | Researching application of AI in products at scale

[Bishal Guha Mallick @bishallick]

@doSwayamExist Do add me, thanks!

[Bishal Guha Mallick @bishallick]

It took very little time for OpenAI to nerf SORA 2 it seems.

Happens every time. Same with Nano Banana from Google.

- Release the most capable model with max compute assigned.

- Get press and viral praise

- Quietly scale back the model’s resources, likely because running it at full capacity is too costly.

- Say absolutely nothing when users start complaining and ride the wave from step 2 as long as possible until the next version is ready for release.

So from an AI builder pov, deploying an AI workflow in production is just the start. To ensure it continues to perform as expected and delivers value, it's crucial to continuously monitor its performance in the real world.

And i think a lot of cloud platforms / providers have built-in tools for this. Like OpenAI’s eval platform and api. For instance, you can schedule monitoring jobs to run at a set frequency, automatically comparing predictions with ground truth labels to track performance degradation over time. These tools can also monitor other critical aspects like token usage, latency, and safety metrics, sending alerts if the model's behavior drifts or violates set thresholds.

[Bishal Guha Mallick @bishallick]

Sooo @OpenAI is turning ChatGPT into a personal OS for your life. Finally your own personal Jarvis?

lol not so soon - i mean not to hype up or anything, but while watching @sama and team speak about Apps SDK at the devday keynote, i kept thinking - wow, so they are really close into turning ChatGPT into a "platform".

Developers can now build native apps that live right inside the chat. Think of AI as a layer between your app and the user - allowing the user to talk to your app/data.

It is a bold move (maybe you saw this coming).

For users, this means apps can pop up when you need them. If you’re talking about a trip, Booking. com might appear. Talk about converting a mockup into a design prototype, and Figma could show up. You can also just call them by name. You can start with an outline and ask Canva to transform it into a slide deck. Or watch a YoutTube video and ask ChatGPT to elaborate on something in the video as you watch. The idea is to make the chat more useful, letting you do things without leaving.

The highlight is that ChatGPT will allow you (as the app publisher) to blend authentication (login), familiar interactive UI elements - like maps, playlists and presentations - into new ways of interacting through conversation.

For developers, this is a HUGE opportunity to unlock a new distribution channel - direct line to ChatGPT's 800 million weekly users. Major players like Spotify and Zillow are already part of the pilot.

And of course - many people are calling it the new "Apple App Store moment".

But it's not that simple. When you build on someone else's platform, you take on their risk. OpenAI controls the rules, the pricing, and the algo that decides which apps get shown (although Sam did say during the keynote that they were planning to launch a directory too).

The bigger issue is the data.

Every app built on ChatGPT helps train OpenAI’s models (maybe not directly). If your app connects housing data, you're teaching ChatGPT how to be a real estate expert. For larger established players, the unique advantage might shrink over time as the platform gets smarter or as a platform they may loose stickiness/redirection etc if users can start completing entire flows within the ChatGPT "ecosystem".

It’s a normal platform dilemma to have I guess. Early developers can win big by getting in early of a massive new channel. But in the long run, the platform itself holds the power. We saw this with Apple and the App Store, or Amazon and Amazon Basics. It plays out slightly differently each time.

But personally, I am really excited.

This isn't just about chat. I think OpenAI is going deeper here - trying to create a new kind of personal OS where AI acts as a layer between you and the services you use. Chat is just ONE of the interfaces. We may see other interfaces like voice, AR (imagine RayBan X OpenAI specs) catching up - these could be interesting experiences.

[Bishal Guha Mallick @bishallick]

@KritikaDarnal @AkasaAir @goibibo @AkasaAir you guys have really poor customer service, and post 6 PM it seems no one picks up the call. 0 support and 0 empathy when a natural calamity happened and people lost their lives. Your team did not even offer to reschedule, simply because they “checked in”

[Kritika Darnal @KritikaDarnal]

Extremely disappointed and heartbroken 💔 with @AkasaAir and @Goibibo.
I missed my #flight from #Bagdogra to Bengaluru on 5th Oct 2025 due to extreme #rainfall and #landslides in the #Darjeeling. Travel was severely disrupted, yet there was no support or help from #AkasaAir.

[Bishal Guha Mallick @bishallick]

PS – check retention, access controls, and PII/redaction settings before you turn it on as per your use case. The logs are not on by default since prompts/completions can have sensitive data. Better to have teams opt-in than accidentally storing everything.

[Bishal Guha Mallick @bishallick]

If you’re switching from OpenAI to Azure OpenAI, check if you will be needing the LLM request/response logs.

We’re used to logs being available on major platforms, but for Azure OpenAI, request/response logs are saved only if you enable them that specific resource.

[Bishal Guha Mallick @bishallick]

Go to “Monitoring” under the specific resource, click Diagnostic settings → + Add diagnostic setting. Pick the categories you need (make sure to include RequestResponseLogs) and choose a destination: Log Analytics/Storage/Event Hub.

[Bishal Guha Mallick @bishallick]

@Elaichaouisamir @OpenAI @OpenAIDevs I think it should be resolved now. Ours reverted to the original amount before the bug hit.

[Bishal Guha Mallick @bishallick]

OpenAI API has gone bonkers! Our credit balance went from $400 to - $3441. Checking the usage data on the platform shows no abnormal usage. It seems like this is a widespread issue. OpenAI, please look into it ASAP.
@OpenAI @OpenAIDevs
More here -
https://t.co/gE7TdCwEbV

[Bishal Guha Mallick @bishallick]

@HugoooDias @OpenAI @OpenAIDevs I don't think it's an UI issue - probably a bug in their credit/billing system. While trying to use the API this error throws up : 'error': {'message': 'You exceeded your current quota, please check your plan and billing details......', 'type': 'insufficient_quota'}

[Bishal Guha Mallick @bishallick]

@y_kurilov @OpenAI @OpenAIDevs The OpenAI team should hopefully restore the credits soon since a lot of folks are facing this issue.

[Bishal Guha Mallick @bishallick]

@mrjamesriley @OpenAI @OpenAIDevs Yes @James - while it definitely takes longer testing the same systems across different service providers to ensure they meet our baseline standards, it pays off in the long term.

[Bishal Guha Mallick @bishallick]

@dharmeshba @SarvamAI If the “text” being send to their TTS API has numbers larger than 4 digits, use commas (e.g., ‘10,000’ instead of ‘10000’). This ensures proper pronunciation as a whole number.

[Bhaskar or Husky @bshubhamb]

Oppenheimer Review (1/n)

It’s not a surprise that I loved the movie. The benefit of watching the first show is a good audience & they maintained deathly silence throughout the entire film. They did quote Bhagavad Gita multiple times in the film - mostly perfectly, except …

[Naman Goyal @NamGoPro]

Hey Twitter Fam,

I recently decided to move on from my Product Manager role at Internshala to explore challenging opportunities.

Please DM or reach out at [email protected] or [phone].

My brief journey below (🧵) -

#job #productmanagement

[Bishal Guha Mallick @bishallick]

@vengurlaa @labnol More like 60-80% in my experience, which is still a good number.

[Bhaskar or Husky @bshubhamb]

18th year away 🌻

Few of us that left home when in primary school have a completely different outlook on life and people.

[Bishal Guha Mallick @bishallick]

Primary skill while user testing: keeping quiet. 😎