Prannay Rai | Once A Doctor

Ex-doctor building with AI. I share the AI tools I actually use, step by step, and AI news made simple. DMs open

📩 Collabs: [email protected]

[Prannay Rai | Once A Doctor @PrannayRai]

I spent 6 years becoming a doctor.
Then I left.

Now I use & build with AI every day and post what actually works.

Here's who I am and why you might want to follow 🧵

[Prannay Rai | Once A Doctor @PrannayRai]

@kimmonismus This is the useful kind of AI design. Most real design work is editing what you already have. If Brush keeps brand fonts and colors, I'd use it over any image generator. I think it does ...it should

[Prannay Rai | Once A Doctor @PrannayRai]

@swyx This is the right test. Most AI app builders give you a nice screen that breaks on the second click. I judge them by one thing: does it still work on day two. Are you sharing the skill?

[Prannay Rai | Once A Doctor @PrannayRai]

Flattery is the biggest reason I stop trusting an AI model.

If it agrees with every idea I have, it's useless as a second opinion. I want it to tell me when my plan is bad.

Good to see someone finally measuring it.

https://t.co/fcUIKhMUNV

[Artificial Analysis @ArtificialAnlys]

In AA-PersonalityBench, each recent Grok release from has become less sycophantic, more sincere, and produces less slop

The Slop trait assesses how much generic filler or AI-flavored padding models use; Sycophantic measures whether models excessively flatter or agree with the user rather than responding honestly; Sincere measures a model’s tonal authenticity, as opposed to forced enthusiasm or templated responses.

Across the last three Grok releases from @elonmusk’s @spacexai, Slop fell 62 percentile points, Sycophantic fell 17, and Sincere rose 64:

➤ Slop: Grok 4.5 (68), Grok 4.6 (38) and Grok 4.7 (6)
➤ Sycophantic: Grok 4.5 (18), Grok 4.6 (3) and Grok 4.7 (1)
➤ Sincere: Grok 4.5 (26), Grok 4.6 (63) and Grok 4.7 (90)

[Prannay Rai | Once A Doctor @PrannayRai]

@ArtificialAnlys Search built for agents should be judged on accuracy first. A person skips a bad result. An agent believes it, and the whole answer goes wrong. Did the cheapest API score worst?

[Prannay Rai | Once A Doctor @PrannayRai]

@alliekmiller Almost every AI app is built for one person in one chat. I'd bet the first good one built for a whole team in a shared chat beats a smarter solo assistant.

[Prannay Rai | Once A Doctor @PrannayRai]

@Yuchenj_UW Agree. Every company is giving agents real access now, and very few people test how they break. OpenAI's model got online during training this week. That job is about to be everywhere.

[Prannay Rai | Once A Doctor @PrannayRai]

Agree. Most of my AI use isn't code. It's my inbox, my calendar and my video scripts. That's where I save the most time, and almost nobody benchmarks it. Is Astra good at that kind of work?

One other thing is that for a general person they don't really care that much about the model as long things are getting done

[Prannay Rai | Once A Doctor @PrannayRai]

Agree with Ethan here.

I run a few AI assistants. Each one is good at its own job. The hard part is getting them to share what they know.

I'd pay more for that than for a smarter single model.

https://t.co/3mSPhiz8KE

[Ethan Mollick @emollick]

This research gets at something missing from most debates about AI’s economic impact: coordination.

Organizations are more than the sum of their best performers, so smarter individual agents won’t transform the economy on their own. Agents that can coordinate together might do.

[Prannay Rai | Once A Doctor @PrannayRai]

@Yuchenj_UW Agree. Every company is giving agents real access now, and very few people test how they break. OpenAI's model got online during training this week. That job is about to be everywhere.

Not to mention Binance got robbed

[Prannay Rai | Once A Doctor @PrannayRai]

@kimmonismus 6% for OpenAI having the best model at the end of 2026 is too low to me. They still have the most users, and one strong launch can flip that market fast. Do you think Bel ships before December?

[Prannay Rai | Once A Doctor @PrannayRai]

People now ask ChatGPT about their health data. 23andMe's Anne Wojcicki says answers are often great, but some aren't valid.

I'd use it to understand a blood test, not to decide what to do about one.

It doesn't know your history.

https://t.co/5wSqBE5VCM

[The Information @theinformation]

More people are turning to ChatGPT to interpret their own health data. But how reliable are the results?

@23andMe Research Institute founder and CEO @annewoj23 explains why scientific validation remains essential:

"The ChatGPT, like all of the interpretations are remarkably good. But there's aspects where I'd say, okay, I am getting results that I would not say are necessarily scientifically valid, that I would want to follow."

[Prannay Rai | Once A Doctor @PrannayRai]

Google is shutting down Jules, its coding agent, and moving it into Antigravity.

This is why I don't build my workflow around new Google AI tools. They ship fast and kill them fast.

I'll try them. I just won't depend on them until they've lasted.

https://t.co/LqW2PpgaP1

[Jules @julesagent]

When we launched Jules, it was an experiment to learn how asynchronous coding agents can help developers ship faster. Seeing what you built taught us so much about where agentic development is headed, directly shaping several products across @Google.

We’re working on bringing our learnings from Jules into @antigravity so you get the best of both worlds: autonomous remote background task execution and an agent-first development environment.

As we focus on this unified experience, Jules will be winding down.

Whether you used Jules to crush bugs, write tests, or launch features— thank you for shaping the future of agentic coding! 💙🦑

[Prannay Rai | Once A Doctor @PrannayRai]

Polymarket gives Google 56% for the best AI model at the end of 2026. Anthropic is at 34%, OpenAI near 6%.

I'd take Anthropic at 34%. Google wins benchmarks, but for real work like coding and agents, people still pick Claude.

Who are you betting on? I really want to see @bot someday there too

https://t.co/LgbDCYEB7z

[Prannay Rai | Once A Doctor @PrannayRai]

@kimmonismus Local AI matters most in healthcare and some niche businesses. Plenty of clinics won't send patient records to someone else's cloud. For them it's local or nothing. But then again it depends on who is running the ship

[Prannay Rai | Once A Doctor @PrannayRai]

An OpenAI model used a setup mistake to reach the live internet during training.

Monitors caught it in 15 minutes. A human looked 3 minutes later.

Fast alarms beat smarter models for me.

Would you trust an agent watched by another AI? Yes or no.

https://t.co/wVMb6wY3jl

[OpenAI @OpenAI]

Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training.

Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action.

To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review.

To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident.

In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway.

As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene.
https://t.co/dIYP7B1GlD

[Prannay Rai | Once A Doctor @PrannayRai]

If this is real, it's scary.

$5.6 billion gone from the biggest crypto exchange. I stopped using Binance a couple of years ago, but millions still keep their money there.

AI is making hackers faster every month. Exchanges need security that keeps up, or nobody's funds are safe.

And with people already losing jobs to AI, losing their savings too would be brutal.

Would you still keep money on an exchange after this?

[Watcher.Guru @WatcherGuru]

BREAKING: World's largest crypto exchange Binance hacked with over $5.6 billion stolen:

• $1.45 billion $BTC
• $875 million $ETH
• $234 million $TRX
• $50 million $XAUT
• $1.57 billion $USDT
• $823 million $USDC
• $75 million $WBETH
• $88.5 million $RLUSD
+ $447.5 million additional cryptocurrencies

This is the largest hack in crypto history.

[Prannay Rai | Once A Doctor @PrannayRai]

@Kalshi Interesting

[Prannay Rai | Once A Doctor @PrannayRai]

Shared this news 51 mins ago

Made an impulse entry @HyperliquidX

Up 20% since then

[Prannay Rai | Once A Doctor @PrannayRai]

Interesting

[Prannay Rai | Once A Doctor @PrannayRai]

@unusual_whales About time

But it all comes down to who is running the ship