Latest / AI Fire Daily / #455 Neil: Gemini 3.2 Pro Leak Just Burned The Current AI Rankings
Transcript
- 0:00Well, we are just five days out from Google I
- 0:02.O., and honestly, the metaphorical dam hasn't
- 0:05just broken today. Oh, it is totally washed away.
- 0:07I mean, nobody saw this massive wave coming.
- 0:09Yeah, it really came out of nowhere. Over the
- 0:11last 48 hours, leaks just completely exploded.
- 0:14We are seeing sudden Gemini 3 .2 rumors everywhere.
- 0:19GPT 5 .6 internal tests are suddenly popping
- 0:22up. Mythos version 2 just successfully completed
- 0:25complex network breaches, and Codex Mobile screenshots
- 0:29are randomly surfacing online. It is wild. Welcome
- 0:33to the Deep Dive. I'm very glad you are joining
- 0:35us today. Today is Thursday, May 14th, 2026.
- 0:39Right. And Google I .O. kicks off in exactly
- 0:41five days. Exactly. So we are taking your sources
- 0:43to examine all this today. We have four massive
- 0:46leaks completely shaking the space. And we really
- 0:48need to look at what this acceleration means.
- 0:51Because this directly impacts your daily workflow
- 0:53and security setups. Okay, let's unpack this.
- 0:56The AI race just got incredibly hot again. It
- 0:59seemed to happen entirely overnight. Google hasn't
- 1:02officially announced Gemini 3 .2 at all. No blog
- 1:05posts, no documentation. Right, zero API changelogs.
- 1:08But the developer community is completely losing
- 1:11its mind anyway. Because developers keep sharing
- 1:14these leaked test results. So let us start with
- 1:17Google's unannounced Pro models. The earliest
- 1:20leaked outputs focus really heavily on SVG generation.
- 1:24Yeah, they do. It is a very specific technical
- 1:27capability to highlight. And for those wondering,
- 1:29an SVG is an image made of math, not pixels,
- 1:33staying sharp at any size. That is exactly why
- 1:35this specific detail matters so much. An AI generating
- 1:39pixel art is basically painting blindly. Right.
- 1:41But an AI generating an SVG is actively writing
- 1:44code. It has to perfectly plot mathematical coordinates
- 1:47and curves. The first test showed a PS5 controller
- 1:49design. The result looked... Well, okay, but
- 1:53nothing truly mind -blowing. Yeah, the buttons
- 1:55did not line up properly at all. The overall
- 1:57shape just felt a bit off ergonomically. It understood
- 2:00the basic geometry, but not human hands. Exactly.
- 2:03But the SVG quality improved significantly in
- 2:05later tests. Another test showed a pelican riding
- 2:08a bicycle. And that specific result looked much
- 2:10more coherent visually. The bird had a very clear,
- 2:13realistic body shape. It featured far better
- 2:16physical proportions. It really understood the
- 2:19abstract relationship between bird and machine.
- 2:21The bicycle itself was mostly complete this time
- 2:24too. No broken parts or completely disconnected
- 2:26wheels. But the really interesting part was the
- 2:29interactive interface. The model generated a
- 2:32full UI right alongside the graphic. Oh, wow.
- 2:35Yeah, users could actively customize the graphic
- 2:37and inspect curve. They could export files and
- 2:40edit specific geometric parts. This suggests
- 2:43a huge shift toward interactive vector workflows.
- 2:46It is not just about basic static image generation
- 2:49anymore. The SPG engine inside Gemini looks incredibly
- 2:53robust now. The coordinate shakes are cleaner
- 2:55and the layouts feel stable. But the UI generation
- 2:57still looks very weak, right? Unfortunately,
- 2:59yeah. The interfaces feel extremely flat and
- 3:02quite generic. They lack the deep polish of a
- 3:05flagship model. One example, use a Web Audio
- 3:07API specifically. It also incorporated Tailwind
- 3:10CSS into the visual design. Yet the final UI
- 3:14still looked very obviously AI generated. Having
- 3:17a great SVG engine but weak UI generation is
- 3:20like putting a world -class engine in a car with
- 3:23no steering wheel. Exactly, or a massive engine
- 3:26bolted to a skateboard. The raw computing power
- 3:29is definitely under the hood. But if the interactive
- 3:32interface remains that clunky... Everyday developers
- 3:35cannot steer it into their actual workflows.
- 3:38Well, part of this issue could just stem from
- 3:40the prompts. If testers focused purely on the
- 3:42vector math quality, the interface results would
- 3:46naturally look much weaker. That makes sense.
- 3:48Based on the leaks, Google clearly prioritized
- 3:51mathematical code quality. User experience definitely
- 3:54looks like the weak spot right now. But the Gemini
- 3:56Pro model surprised people in other ways, too.
- 3:59Another leak showed its true creative generation
- 4:02power. Right. From one simple prompt, it created
- 4:04four distinct robots. And all four robot designs
- 4:08look surprisingly good. They were not just lazy,
- 4:11simple copies of each other. They had completely
- 4:13different colors and structural shapes. Yeah.
- 4:15Unique, small design details were built right
- 4:17in. This proves it can generate multiple creative
- 4:20variations accurately. It does not just retune
- 4:22one repetitive, hallucinated output. This is
- 4:25incredibly useful for vector asset design workflows.
- 4:29Icon designers and mascot creators are going
- 4:31to absolutely love this. But the flash variants
- 4:34are where things truly heat up. Let's talk about
- 4:37that. There are multiple flash variants in active
- 4:40internal testing. The leaks mention code names
- 4:43Fanta, Sprite, and Cola. These are internal names
- 4:47for different experimental model configurations.
- 4:50Each variant targets completely different performance
- 4:52optimization goals. One might be tuned for pure
- 4:55processing speed. Another might focus purely
- 4:57on high factual coding accuracy. Right, this
- 5:00is standard operating practice at major AI labs.
- 5:03They train several versions and basically pick
- 5:05the best performer. And the Fanta variant specifically
- 5:08showed very competitive benchmark results. It
- 5:10is testing incredibly well against the Pro model.
- 5:13Fanta is nearly rivaling Pro in many complex
- 5:15aspects. Which raises a big question. Will the
- 5:18Fanta Flash variant cannibalize the Pro model's
- 5:20market if speed wins out? If Google ships a Flash
- 5:23variant this strong, It changes everything. Users
- 5:26prioritize speed for fast, practical, everyday
- 5:29engineering work. It could completely rewrite
- 5:32current pricing and tier models. So speed and
- 5:35daily efficiency might actually win out over
- 5:38raw, complex reasoning power. Right. And that
- 5:40is exactly what we are seeing. Pro handles complex
- 5:43reasoning and multi -step architectural planning
- 5:46tasks. While Flash handles the fast, repetitive
- 5:49practical daily coding work. Exactly. But here's
- 5:53the really critical thing to consider. If Google
- 5:56nails that daily efficiency with their models,
- 5:59OpenAI cannot just sit around on their hands
- 6:01quietly. Very true. What's fascinating here is
- 6:03the sheer development speed. The leaks show OpenAI
- 6:06is definitely feeling the heat. While everyone
- 6:09watches Google, OpenAI is testing very quietly.
- 6:12GPT 5 .6 is already in active internal testing.
- 6:15Yeah, the first development checkpoints were
- 6:16evaluated very recently to internal code. code
- 6:18names showed up in the recent leaf. Amber alpha
- 6:20and beacon alpha, right? Those are the ones.
- 6:23These are likely two advanced variants being
- 6:25heavily tested. They want to see which gives
- 6:27better complex benchmark results. They run advanced
- 6:30candidates in parallel and pick one. We might
- 6:33see these internal candidates appear. very quietly
- 6:36soon. They often show up on LM Arena testing
- 6:38first. They just run public blind testing without
- 6:40telling anyone beforehand. The gap between major
- 6:43releases is shrinking drastically. I mean, GPT
- 6:475 .5 just launched a few short weeks ago. And
- 6:49it was a massive structural success for the company.
- 6:52Reviewers said it outperformed Anthropix Opus
- 6:54easily. That momentum gives OpenAI strong reason
- 6:58to push forward. especially with Google about
- 7:00to show new flagship models. If OpenAI waits
- 7:03too long, they lose the narrative entirely. Some
- 7:07reliable sources actually suggest a release by
- 7:09next month. Wow. That would be one of the shortest
- 7:11release gaps ever. One major reason for this
- 7:14acceleration is quite profound. AI itself now
- 7:17actively helps build new AI models. Right. Models
- 7:19constantly assist with complex code generation
- 7:22tasks now. They handle vast data preparation
- 7:25and evaluation during training. They optimize
- 7:27the actual architectural training process itself
- 7:30continuously. They completely remove the human
- 7:33bottleneck. from data collection. The model generates
- 7:36millions of complex edge case coding problems.
- 7:39And then it autonomously solves those exact same
- 7:41coding problems. And finally, an internal reward
- 7:45model grades the final solution. You know, Ibeat,
- 7:48it is a very humbling reality. I still wrestle
- 7:51with prompt drift myself on a daily basis. Yet
- 7:54these models are already debugging their own
- 7:56training code. It really makes you realize how
- 7:58quickly we become bottlenecks. Models generate
- 8:01synthetic data way more efficiently than human
- 8:04teams. They instantly spot critical processing
- 8:06bottlenecks earlier than us. Every lab knows
- 8:09slowing down means losing crucial ground. Absolutely.
- 8:12So since AI is building AI right now, are we
- 8:15going to see massive capability leaps or just
- 8:17better integration? We might not see those massive
- 8:19historical capability jumps anymore. Early years
- 8:22were defined by those massive intelligence leaps.
- 8:25Now, updates will focus heavily on pure processing
- 8:28speed. They will focus on smoother workflows
- 8:30and deep personalization. Right. Smoother workflows
- 8:33and integration matter far more now than chasing
- 8:36massive capability leaps. And that concept of
- 8:39smoother workflows is actually very important.
- 8:41It connects directly to highly complex, high
- 8:45-stakes digital environments. Which brings us
- 8:47to the intense cybersecurity sector today. This
- 8:50specific space does not get enough mainstream
- 8:53attention at all. It really doesn't. And the
- 8:56cybersecurity AI crown is shifting incredibly
- 8:58fast lately. We started with Claude Opus 4 .6,
- 9:02leading everyone. Then the original Mythos preview
- 9:04appeared on the scene. It stayed slightly behind
- 9:07the Opus model initially. But then GPT 5 .5 Cyber
- 9:10closed that performance gap. It was tuned specifically
- 9:14for complex security network workflows. Right,
- 9:17and now Mythos Preview Version 2 is officially
- 9:19arrived. Some elite developers are currently
- 9:21pronouncing it as Mythos. It just received a
- 9:24major upgrade that is turning industry heads.
- 9:27Early benchmark results show a clear capability
- 9:29jump immediately. Institutions like the AI Security
- 9:33Institute are actively testing it. It has confidently
- 9:36jumped back into the absolute lead. This constant
- 9:39back and forth really shows the development speed.
- 9:42The competitive cybersecurity AI race is incredibly
- 9:46intense right now. The most striking test involves
- 9:49a complex network simulation. They use a 32 -step
- 9:52corporate network attack simulation. And this
- 9:55is definitely not a simple theoretical toy benchmark.
- 9:58The model needs to identify obscure digital system
- 10:01vulnerabilities. It has to actually chain multiple
- 10:04obscure software exploits together. It must execute
- 10:08lateral movement smoothly through the whole network.
- 10:11moving from a compromised printer straight to
- 10:13an admin laptop. It has to complete a full, complex
- 10:17attack path. A human cybersecurity expert needs
- 10:19roughly 20 hours total for this. But this new
- 10:22mythos preview completed the entire simulation
- 10:25rapidly. Yeah, it took only 6 to 10 digital attempts
- 10:27total. This is a dramatic difference in operational
- 10:30security capability. Whoa, Pete. Imagine simulating
- 10:33a 32 -step network breach in just a few tries.
- 10:36The magnitude of that technical achievement is
- 10:38honestly staggering. A 32 -step attack chain
- 10:41requires deep architectural understanding. It
- 10:44must know exactly which obscure exploits actually
- 10:47work. It strings everything together in the correct
- 10:50logical sequence. The fact that an AI does this
- 10:52reliably really matters. It is a massive structural
- 10:55deal for the entire industry. But does giving
- 10:58an AI the ability to chain exploits hand a dangerous
- 11:02weapon straight to bad actors? That is where
- 11:05things become significantly more complicated
- 11:07ethically. The exact same model defending a network
- 11:10can attack it. The potential risks of widespread
- 11:13misuse are incredibly high. That is exactly why
- 11:16software releases are carefully staged. Governments
- 11:19and security institutions usually get private
- 11:21access first. They have crucial time to prepare
- 11:24necessary security defenses. They actively do
- 11:26this before tools become widely available. Exactly.
- 11:29Defense and attack grow together. Early access
- 11:31for human defenders is absolutely crucial. But
- 11:33this immense autonomous computing power is expanding
- 11:36very rapidly elsewhere, too. It is not staying
- 11:39securely locked inside corporate data centers
- 11:41forever. Right. It is moving directly to your
- 11:43personal mobile phone. recent leaked screenshots
- 11:46started showing OpenAI's codecs running locally.
- 11:49It is operating directly from a standard mobile
- 11:51device. This immediately caught the attention
- 11:54of many online developers. It does not look like
- 11:57a highly polished app yet. It seems much closer
- 12:00to a remote coding workflow. Yeah, they are using
- 12:02a mobile CLI right now. Let's quickly define
- 12:05that. A mobile CLI is a text -based command line
- 12:08to control remote computers from your phone.
- 12:11That is a perfect, concise summary of the tool.
- 12:14You can manage complex coding tasks away from
- 12:16your desk. You can comfortably review outputs
- 12:19and interact directly with codecs. Earlier leaks
- 12:22hinted at much deeper workflow integrations coming
- 12:24too. Tools like NotionHQ are heavily rumored
- 12:27right here. You could plan, document, and code
- 12:30in one single place. There is also specific mention
- 12:33of remote control features. You can trigger heavy
- 12:35codecs compiling tasks directly from your phone.
- 12:38Because the heavy processing still runs safely
- 12:40on a remote server. Rohan Burma from the Core
- 12:43Codex team was actually mentioned. He was specifically
- 12:46cited as a primary source for this leak. Which
- 12:48adds major industry credibility to these current
- 12:51screenshots. This definitely does not look like
- 12:53random internet speculation. If Codex arrives
- 12:55on mobile, it completely changes developer workflows.
- 12:59Right now, most professional tools need a heavy
- 13:01desktop IDE. Mobile access lets you actively
- 13:04review code from anywhere. You can approve massive
- 13:07pull requests while on the go. It's not about
- 13:09typing syntax with your thumbs. It's like carrying
- 13:12a senior developer in your pocket as a remote
- 13:14control. That is exactly how it will psychologically
- 13:17function for people. You could be sitting on
- 13:19a busy train commute. Or you could be sitting
- 13:21quietly in a long meeting. You can still keep
- 13:24your complex coding projects moving forward.
- 13:26You open your phone and check the project documentation.
- 13:30You trigger Codex to actively review the new
- 13:33logic. Codex finds a bug, fixes it. and pushes
- 13:36the commits. Will developers actually want to
- 13:39manage heavy processing tasks away from their
- 13:41desks? They'll definitely not do heavy typing
- 13:43on a mobile screen, but they will urgently want
- 13:46to unblock their teammates. They will want to
- 13:47review AI -generated code snippets very quickly.
- 13:50No, it adds flexibility to keep projects moving
- 13:53anywhere without replacing the desktop entirely.
- 13:57Mid -roll sponsor break. So what does this all
- 13:59mean? If we connect this to the bigger picture
- 14:01today... A very clear, undeniable historical
- 14:05pattern is showing up. All four leak stories
- 14:08point to the exact same trend. AI development
- 14:10labs are releasing massive models faster than
- 14:13ever. The technical time between major versions
- 14:15is completely collapsing now. Internal testing
- 14:18cycles overlap heavily with public release cycles.
- 14:21Significant leaks happen before official corporate
- 14:24announcements way more often. Two massive market
- 14:27forces constantly drive this rapid acceleration.
- 14:30Advanced AI tools actively make software development
- 14:33much faster. An extreme competitive corporate
- 14:36pressure drives the entire industry forward.
- 14:39No single lab can afford to look completely behind.
- 14:41We are looking at a massive volatile pressure
- 14:44cooker. OpenAI aggressively pushes 5 .6 because
- 14:48of Google I .O. Google rushes Gemini development
- 14:50because Anthropic's Mythos is gaining ground.
- 14:53Anthropic accelerates Mythos testing because
- 14:55of GPT -Cyber's incredible performance. Every
- 14:59strategic move by one lab creates intense market
- 15:01pressure. It forces the other major competitors
- 15:03to sprint even faster. That is exactly why we
- 15:06see this massive flood of leaks. The broader
- 15:09AI space moves way faster than people can track.
- 15:12The real practical lesson here is about managing
- 15:14the pacing. The technical field heavily rewards
- 15:18people who stay consistently informed. You must
- 15:21actively act on what actually matters most. Right.
- 15:24Do not treat every single internet leak as gospel.
- 15:27You should patiently wait for the official software
- 15:30releases. Compare the real technical outputs
- 15:32when they are finally public. Pay close, careful
- 15:35attention to actual real -world software deployments.
- 15:38Track the quiet, unannounced deployments on LM
- 15:41Arena closely. The most meaningful software improvements
- 15:44are often very quiet. They just secretly make
- 15:47your daily work much faster. They smoothly integrate
- 15:50into your workflow without you even noticing.
- 15:52Two secs silence. Here's where it gets really
- 15:54interesting for the future. Google I .O. officially
- 15:57happens on May the 19th. You need to keep a very
- 15:59close eye on it. We must actively see if the
- 16:03pro versus flash reality actually matches the
- 16:06leaks. We desperately need to see if flash truly
- 16:08wins out. I want to leave you with one final
- 16:10thought today. We know AI is successfully debugging
- 16:13its own training code. It is autonomously accelerating
- 16:16the software development loop drastically. What
- 16:19happens when these models begin designing the
- 16:21next generation of physical hardware and chips
- 16:23optimized purely for themselves? That is a truly
- 16:26staggering philosophical concept to consider.
- 16:29They could autonomously design the next generation
- 16:31of silicon chips. They could completely remove
- 16:34the human bottleneck from the physical timeline.
- 16:37It is something to seriously ponder on your own.
- 16:39Stay sharp.