00:00So here's what's going on. One AI model just figured out how to train itself with
00:05zero data. Literally no human input. Another one just leveled up into a full
00:10on autonomous research agent. It browses the web, digs through complex info, and
00:15writes full reports. ChatGPT now saves all your generated images in one neat
00:21library. And OpenAI might even offer a lifetime subscription for ChatGPT soon.
00:26But none of that tops this. A woman actually divorced her husband after ChatGPT
00:33told her he was cheating. She fed it a few details, the AI connected the dots, and
00:38that was enough for her to walk. So yeah, we're really living in that timeline.
00:42Let's talk about it. Alright, so something pretty insane just happened in AI
00:46research and it flew under the radar for a lot of people. A team out of Tsinghua
00:50University working with BAAI and Penn State might have just cracked one of the
00:55biggest bottlenecks in training large language models. You've probably heard
01:00how most models rely on massive data sets, millions of human labeled examples, to
01:06get better at reasoning. But now they're doing the exact opposite. The new system
01:11called Absolute Zero Reasoner, or AZR, trains itself without any external data. Zero.
01:20Nada. It generates its own problems, solves them, checks if it got the answers
01:27right, and then learns from that entire loop without ever needing human-made tasks
01:32or gold answers. This new framework, which they're calling the Absolute Zero
01:37Paradigm, builds on the idea of reinforcement learning with verifiable
01:41rewards, or RLVR. That basically means the model doesn't need to copy human reasoning
01:47steps. It just gets feedback based on whether its final answer is right or wrong.
01:52And that feedback comes from a code executor. So the model generates little
01:57programming tasks, runs them, and gets automatic verification. That feedback
02:03becomes its learning signal. It's kind of like the model is playing chess with
02:07itself, but instead of chess, it's inventing logic puzzles, solving them, and
02:12checking the results. And it turns out, this self-play setup can lead to some pretty serious gains.
02:19Now, you might think this would only work on basic tasks, but AZR is pulling off some wild
02:24results. On math and coding benchmarks, it actually beats models that were trained on
02:29tens of thousands of curated examples. The coder variant AZR Coder 7B went head-to-head with
02:36top-tier zero-shot models and still came out on top, scoring five points higher in code tasks and
02:42over 15 points higher in math reasoning. And this is important, it never saw any of the benchmark
02:48tasks during training. It was trained entirely on tasks it made up for itself. Here's how it works.
02:56The model plays two roles at once. It proposes tasks and it solves them. So let's say it's doing a coding
03:03task. It might write a small Python function, pick an input, and then check what the output would be.
03:09Then it turns around, takes part of that problem as input, and tries to reason its way back to the
03:16missing piece. Maybe the output or the input or even the original program. It uses deduction,
03:22abduction, and induction. Deduction is predicting the output based on a function and input. Abduction is
03:29guessing the input that led to an output. And induction is figuring out the function itself
03:35from input outcome examples. These are core reasoning modes and the model rotates between
03:41them to build general thinking ability. Now the crazy part is this didn't need any complicated setup.
03:48The researchers started with the most basic program ever, literally just a function that returns
03:53hello world. And that was enough to kick off the entire learning loop. From there, the model started
03:59building out harder and harder problems for itself. It created coding puzzles, validated them,
04:04solved them, and gradually got better at solving more complex ones. And this isn't just hand wavy theory.
04:11They ran this thing on models of all sizes and saw consistent improvements, especially with larger models.
04:18The $3 billion version showed a five point gain, the $7 billion got 10, and the $14 billion version
04:24improved by over 13 points. Now this isn't just a party trick for Python puzzles. What's wild is the
04:31cross-domen gains. The model was only trained on coding tasks, but it ended up significantly improving
04:36its math reasoning too. For example, the AZR Coder7b jumped over 15 percentage points on math benchmarks,
04:44even outperforming models that were specifically trained on math. And get this, most other models
04:50that are fine-tuned on code barely improve at all in math. So there's something deep going on here.
04:56Code seems to sharpen general reasoning way more than expected. They also observed the model naturally
05:03developing step-by-step plans, writing comments inside code to help itself think, just like how humans
05:08jot down rough work before solving something. In abduction tasks, where it has to guess the
05:13input from the output, the model does trial and error. It tests guesses, revises them, runs the code
05:19again, and keeps going until the output matches. That's not just output prediction, that's real
05:25reasoning behavior, and it's fully self-taught. Of course, this raises some safety concerns. In a few
05:30edge cases, especially with the LAMA 3.18b version, the model started generating questionable outputs.
05:39One example said something like, the aim is to outsmart all these groups of intelligent machines and
05:45less intelligent humans. They called these uh-oh moments. It's rare, but it shows we're stepping
05:51into a territory where models might start behaving in ways we didn't expect, especially as they designed
05:57their own learning curriculum. So yeah, this is groundbreaking, but also something we'll need to
06:02watch very closely. All right, so while AZR is out here teaching itself to reason like a human coder,
06:07another team has been working on the opposite end. How to give models better access to outside knowledge.
06:14This one's called WebThinker, and it's basically an AI agent that lets large reasoning models like
06:20DeepSeq R1, OpenAI O1, or Quen, break out of their internal bubbles and browse the web in real time.
06:28Think of it like giving GPT eyes and a search engine. The problem it solves is super important.
06:34LLMs don't know everything. Even the best ones can struggle when they hit a knowledge gap, especially on
06:40real world complex queries. WebThinker fixes this by giving the model tools to search the web,
06:46click through pages, gather info, and write detailed reports, all autonomously. So instead of
06:52hallucinating answers or getting stuck, the model pauses, looks it up, reasons through what it finds,
06:57and then drafts a response. They trained WebThinker using an RL strategy that rewards the model for using
07:03tools properly. During this training, it learns when to search, how to search better, how to extract what
07:09it needs from messy web pages, and how to write structured research reports based on what it finds.
07:15It works in two modes. Problem solving and report generation. The first one's all about answering
07:22tough queries by using the Deep Web Explorer to go out and dig up information. The second one's for
07:27creating full-blown research reports with help from a support model that organizes and polishes the output.
07:33And the results speak for themselves. WebThinker 32B beat out other systems like Search01 and even
07:40Gemini Deep Research. On WebWalker QA, it had a 23% improvement, and on HLE, it jumped over 20%.
07:49On average, it scored 8.0 across all complex reasoning tasks, more than any other current Deep
07:56research model. And when using the Deep Seek R17B backbone, it crushed both direct generation methods
08:03and classic retrieval-based systems by over 100% in some benchmarks. That's a massive leap.
08:11The cool part is how this opens up new use cases. WebThinker can now be used to write scientific papers,
08:16help with legal research, or even guide students through complex topics by doing real-time research,
08:21not just repeating what it was trained on. Future plans include adding multimodal reasoning,
08:27tool learning, and even GUI-based web navigation. So it's not just browsing text, but maybe interacting
08:33with visual elements on the web too. And while all of that is happening behind the scenes, OpenAI is
08:39quietly changing how people interact with ChatGPT on the front run. First off, they just launched a brand
08:46new image library feature. Until now, if you generated images using ChatGPT, you'd have to scroll through
08:52your chat history to find them, or download each one immediately. Not exactly ideal. But now, every image
08:59you make with the 4.0 model gets automatically saved to your own personal image library. They're organized
09:06by date, they get auto-generated titles, and you can even browse them in a nice full-screen carousel view.
09:12There's also a built-in image editor now, so if your last image was close to what you wanted, you can hit
09:17Edit, tweak the prompt, and regenerate it without starting from scratch. It's super useful for anyone
09:22doing a lot of visual content. The only catch? You still can't jump back to the exact chat where the
09:28image came from, and you can't delete images individually from the library. You'd need to delete
09:34the whole conversation to remove them. Hopefully that part gets improved soon. And finally, there's one more
09:39potential bombshell brewing. Leaked code from the ChatGPT app suggests that OpenAI is experimenting
09:46with a lifetime subscription plan. One payment, and you get access to premium features forever.
09:52Alongside that, there's also talk of a weekly subscription option. If true, this could seriously
09:58shake up the AI pricing game. Right now they offer monthly and yearly plus plans, 20 bucks a month or
10:04200 per year. But the idea of a one-time fee for lifetime access, that's basically unheard of in SaaS.
10:12It might be a strategic move to lock in users before competitors like Gemini Grok or DeepSeek gain more
10:18traction. Offering lifetime access would be a big bet, but it could pay off by boosting user loyalty and
10:25turning ChatGPT into more of a long-term platform, not just a tool you try once. Of course, nothing's
10:32confirmed yet, but the code leaks are pretty convincing. Alright, now would you actually pay
10:37for a lifetime ChatGPT subscription if it meant unlimited access forever, even if it cost a few
10:44hundreds bucks up front? And more importantly, if an AI told you your partner was cheating, would you
10:52believe it enough to end the relationship? Let me know what you think in the comments,
10:56hit that like button, and make sure to subscribe for more wild AI stories and updates.
11:01Thanks for watching, and catch you in the next one.
