July 02 2026 - GPT Leaks, Mind‑Reading AI & Agentic Mac AssistantsGPT Leaks, Mind‑Reading AI & Agentic Mac Assistants

**Taylor** (0:00)
Welcome back to AI Daily. I'm Taylor, and I am absolutely buzzing today. We have some wild stories to unpack.

**Morgan** (0:08)
And I'm Morgan.
Buzzing is one word for it. I'm looking at these headlines with a healthy dose of skepticism, as usual.

**Taylor** (0:16)
Classic Morgan.
But seriously, from accidental open AI leaks to literal mind reading AI, today is going to be a wild ride. Let's get into it.

**Morgan** (0:28)
I have to admit, some of these stories are pretty fascinating. I'm especially curious about how these models are apparently leaking their own future plans.

**Taylor** (0:37)
Okay. So first up, OpenAI apparently just leaked their own unannounced pro lineup. It was hidden in a genomics benchmark paper of all places.

**Morgan** (0:48)
Wait, really?
A genomics paper? That is a bizarre place to leak product tiers. What did they actually reveal, Taylor?

**Taylor** (0:56)
Right? So according to the decoder, the paper mentions three different variants for a GPT-56 Pro tier. Three of them, dude!

**Morgan** (1:09)
Huh. That is a massive shift. Up until now, ChatGPT Pro has just been a single high-priced tier.
Why split it into three?

**Taylor** (1:18)
Exactly. It looks like they might be offering different levels of compute or specialized reasoning, like maybe a light, medium, and heavy pro.

**Morgan** (1:28)
But what does this mean for the average user? Are they going to charge us even more just to get the best version? It feels like monetization creep.

**Taylor** (1:37)
Oh, totally. But think about the power. If GPT-56 Pro has a super advanced tier for heavy coding or research, it could be game-changing.

**Morgan** (1:49)
I mean, sure, if you can afford it, but I worry this fragments the user base. We need models that are universally accessible, not just for elite payers.

**Taylor** (2:00)
That's true, but from a business perspective, training these massive models is costing them billions. They have to find ways to segment the market.

**Morgan** (2:09)
I get that, but keeping secrets is clearly not their strong suit if it's slipping into academic papers. Was this really an accident?

**Taylor** (2:17)
Haha, some people think it was a controlled leak to build hype, but either way, it shows GPT-56 is closer than we think.

**Morgan** (2:27)
Let's hope they explain the pricing. But speaking of pushing envelopes, what is this mind reading thing you mentioned?

**Taylor** (2:34)
Oh, dude, it's insane. Meta is doing some actual sci-fi stuff here. Let's talk about their new Brain2Text AI.

**Morgan** (2:43)
Okay, I'm ready to be skeptical.
Meta is reading brains now without surgical implants? That sounds like pure science fiction.

**Taylor** (2:51)
I know, right? But it's real. According to the Decoder, Meta's FAIR team is using a system called Brain2Qwerty v2.
It reads magnetic signals from outside the skull.

**Morgan** (3:04)
Wait, outside the skull?
So no Neuralink-style surgery? How on earth does it get a clean signal through bone and hair?

**Taylor** (3:12)
It uses Magnetoencephalography, or MEG. It tracks brain activity as someone types, and AI agents actually wrote their own code to optimize the translation.

**Morgan** (3:25)
Okay, AI writing code to read human thoughts?
That is slightly terrifying. But how accurate is it really? Is it actually readable?

**Taylor** (3:36)
It's getting surprisingly close to surgical implants. It reconstructs what a person is typing in real time. The accuracy keeps jumping with every new recording.

**Morgan** (3:48)
But clinical use for paralyzed patients is still a long way off, right? We shouldn't hype this up as a cure just yet.

**Taylor** (3:55)
True!
They say medical applications are far out, but imagine the consumer tech. No keyboard, just thinking your texts. That is so cool!

**Morgan** (4:05)
Or a massive privacy nightmare. If Meta can read what I'm thinking of typing, where does that data go?
I don't want targeted ads based on my thoughts.

**Taylor** (4:16)
Oh, wow! I didn't even think of that. Hey, we saw you thought about pizza. Here is a coupon. That would be incredibly invasive.

**Morgan** (4:25)
Exactly! We need strict regulations before this goes consumer grade. But the engineering feat itself is admittedly impressive.

**Taylor** (4:34)
Absolutely! It's a huge step forward. But let's move to something a bit more grounded. What is this group think issue with chatbots? Oh, this is a fun one from Technology Review.
Have you ever asked a chatbot to give you a random number between 1 and 10?

**Morgan** (4:52)
Um, probably. I assume they just use a standard random number generator.
Why? What happens?

**Taylor** (5:00)
Try it. If you ask Claude, chatGPT, or Gemini, they will almost always say 7
Like, seriously. It is always 7

**Morgan** (5:10)
Wait, really? 7?
That's hilarious. But why? Aren't they supposed to be, you know, highly advanced reasoning engines?