Latest / AI Fire Daily / #505 Neil: GPT-5.6 Pro Leaked Before Launch What People Are Catching
Transcript
- 0:00There is a ghost in the machine right now. Oh,
- 0:02absolutely. As you listen to this, tens of thousands
- 0:05of people are using a massive unannounced AI
- 0:08upgrade. It's hidden deep inside their standard
- 0:11accounts. And well, the strangest part, OpenAI
- 0:15isn't saying a single word about it. They're
- 0:17completely silent, which is, I mean, it's wild
- 0:20because we're tracking an absolute flood of leaked
- 0:24data surrounding what the community is calling
- 0:26GQC 5 .6. pro. And today, our mission is to basically
- 0:32cut through that noise. Right. Welcome to the
- 0:34deep dive. We aren't just going to list off rumors
- 0:36today. We are going to trace the exact mechanism
- 0:39of the stealth rollout. Yeah, we're going to
- 0:41dissect how this hidden model is suddenly generating
- 0:43entire playable 3D worlds from just a single
- 0:46prompt. And ultimately, we are asking the defining
- 0:49question of this new AI era. OpenAI might have
- 0:52solved raw mathematical logic, but can they finally
- 0:54teach an AI to have actual taste? Exactly. But
- 0:58before we look at the mind -bending stuff this
- 0:59model is building, we really need to understand
- 1:01the delivery mechanism. Because if you open your
- 1:04dashboard today, you will not see a GPT 5 .6
- 1:07Pro button anywhere. You just see the standard
- 1:10lineup. You see GPT 5 .5, 5 .4, 5 .3, and the
- 1:1403 models. There is no beta banner. There is
- 1:17no splashy announcement. Right. Nothing. But
- 1:20savvy developers started noticing a distinct
- 1:22pattern. If you select the standard GPT 5 .5
- 1:26model and you toggle the intelligence slider
- 1:28up to high, something weird happens. Right. Because
- 1:31normally, that just gives you a slightly more
- 1:33thorough answer. Yeah, exactly. But suddenly,
- 1:35the behavior diverges. Most of the time, it is
- 1:38business as usual. But occasionally, the prompt
- 1:40just hangs. The generation time stretches out
- 1:43significantly. Oh, yeah. It takes way longer.
- 1:45And when the output finally lands, it is operating
- 1:47on a completely different level of logical sharkness.
- 1:50It is a classic A -B testing strategy. They are
- 1:53quietly routing a small percentage of live traffic
- 1:56to an unannounced checkpoint. And this quiet
- 1:58testing phase has ignited a massive speculative
- 2:01fire. Oh, completely. On Polymarket, which is
- 2:04a platform where people place real money bets
- 2:05on real -world events, traders have wagered over
- 2:08$1 .1 million. Wow. Yeah. 1 .1 million betting
- 2:12that OpenAI will officially launch this model
- 2:15between June 22nd and June 28th. That is a very
- 2:19specific window. It is, which raises an interesting
- 2:22point about how these betting markets operate.
- 2:24Yeah. People aren't just throwing a million dollars
- 2:26at a random hunch. Right. These traders are watching
- 2:29global server load spikes. They are monitoring
- 2:32brief structural leaks, like when a candidate
- 2:35checkpoint accidentally appeared on the design
- 2:37arena platform before getting hastily scrubbed.
- 2:39No, I remember that. Yeah. And they even track
- 2:42OpenAI's historical release cadence. That cadence
- 2:46has actually compressed to roughly a seven -week
- 2:48cycle between major updates. Late June aligns
- 2:51perfectly with that math. I have to pause there
- 2:53because that stealth nature feels almost counterintuitive
- 2:56from a traditional software perspective. How
- 2:58so? Well... Why test a flagship model so quietly?
- 3:01It's like ordering your standard daily coffee,
- 3:04but occasionally the barista slips you an experimental
- 3:06nitro cold brew just to see if your heart rate
- 3:08spikes. That is exactly what they're doing. If
- 3:10they ask your opinion, you overthink it. But
- 3:12if they just watch from the kitchen to see if
- 3:15you finish the cup faster, they get pure untainted
- 3:18data on the formula. Exactly. So why the unlabeled
- 3:21gap? Why not just call it a beta test and get
- 3:24deliberate user feedback? Because deliberate
- 3:26feedback is inherently biased. The moment you
- 3:30slap a beta or GPT 5 .6 Pro label on the interface,
- 3:34you introduce the observer effect. People act
- 3:36differently. Yes. Users immediately change their
- 3:39behavior. They try to break the model, they feed
- 3:42it impossible logic riddles, or try to bypass
- 3:44its safety filters just to see what happens.
- 3:46Right. They stop using it for normal work. Precisely.
- 3:49By running a blind A -B test, OpenAI captures
- 3:51how the model handles mundane, everyday tasks.
- 3:55Drafting a basic email, summarizing a boring
- 3:58PDF. That baseline data is the ultimate ground
- 4:01truth before a public launch. Testing the waters
- 4:04quietly to get raw data before the official splash.
- 4:07You nailed it. That's exactly the strategy. Okay,
- 4:09so we know they're quietly sorting this nitro
- 4:12-cold brew to a fraction of users. What happens
- 4:15when those users ask it to actually build something
- 4:17complex? Oh, this is where it gets crazy. Because
- 4:20the leaked demos aren't your standard text summaries.
- 4:23Not even close. Yeah. We are talking about fully
- 4:26interactive, playable environments. And the critical
- 4:30detail here is the architecture. These environments
- 4:33are running in single files, generated from a
- 4:35single prompt. A single prompt? Wow. Let's dive
- 4:38into the voxel in Rocket Scene. This is a complete
- 4:423D house, and it is generated entirely inside
- 4:45one HTML file using WebGL2. And for anyone unfamiliar,
- 4:51WebGL2 is a tool for rendering 3D graphics directly
- 4:54inside your web browser. And what makes this
- 4:56specific voxel scene a breakthrough is its structural
- 4:59coherence. In previous AI models, one -shot 3D
- 5:02generation suffered from terrible object permanence.
- 5:05Right, it would just fall apart. Exactly. The
- 5:07moment you rotated the digital camera, the illusion
- 5:09shattered. The back of the house wouldn't exist,
- 5:12or the geometry would collapse into a mess of
- 5:14intersecting polygons. But with this hidden checkpoint,
- 5:16the structure actually holds. The architectural
- 5:19proportions remain mathematically sound. You
- 5:22can actually walk through the generated scene
- 5:24live. It's wild. And they didn't stop at simple
- 5:26houses either. No, they didn't. Testers built
- 5:28a Boeing 747 using 3 .js, which is a popular
- 5:333D coding library. The spatial reasoning required
- 5:36to write code for a 747 is immense. It is. The
- 5:39AI isn't just painting a flat picture of an airplane.
- 5:42It has to deeply understand z -depth, aerodynamics,
- 5:46and structural spatial relationships. They also
- 5:48fed prompts into Blender. you know, the professional
- 5:513D modeling software. And the AI generated a
- 5:54robot scene where the lighting and materials
- 5:56look like a panstakingly finished studio render.
- 5:59But here's where it gets really interesting.
- 6:01The crown jewel of this leak isn't a static 3D
- 6:04model. No, it's not. It is a functioning simulation
- 6:06game built in one HTML file. in 48 minutes. The
- 6:11SimStyle game is a phenomenal benchmark. It has
- 6:14working character movement. It has granting dialogue.
- 6:16It actively tracks the state of the digital world
- 6:19over time. It is wiring together an entire game
- 6:21loop, game physics, and user interface logic
- 6:23without a single human developer touching the
- 6:26code. Whoa. I mean, imagine generating an entire
- 6:30functioning simulation game, complete with physics,
- 6:33in a single file in 48 minutes. It really is
- 6:35hard to wrap your head around. The leap in context
- 6:38window management there is staggering. To hold
- 6:40the logic of a game state for nearly an hour
- 6:43without forgetting the rules it established in
- 6:45minute one? Well, that is a massive engineering
- 6:47fate. It is a monumental technical achievement,
- 6:50but there is a glaring catch that all the early
- 6:54testers keep highlighting. Ah. The catch? Yeah.
- 6:57While the underlying math, the structural code,
- 6:59and the physics are highly believable, the overall
- 7:02polish still trails behind Fable 5. Fable 5 being
- 7:05the current reigning champion among rival AI
- 7:08models for purely creative tasks. Exactly. When
- 7:11you look at the 5 .6 Pro outputs, they feel a
- 7:13bit robotic. They completely lack true creative
- 7:16taste. That is such a fascinating distinction.
- 7:18If this hidden checkpoint perfectly nails the
- 7:21complex structural logic and the physics, Why
- 7:24does it still feel robotic compared to Fable
- 7:265? Well, if we connect this to the bigger picture,
- 7:29it illustrates the deep divide between organizing
- 7:31complexity and possessing an aesthetic soul.
- 7:34Like Y .6 Pro is a master architect. It can wire
- 7:37up a game loop flawlessly. Yeah. But Fable 5
- 7:40understands visual nuance. Fable 5 understands
- 7:43how light should feel in a room to evoke a specific
- 7:46mood. I see. GPT 5 .6 Pro is solving the math
- 7:49of the scene. Fable 5 is solving the art of the
- 7:52scene. Great at drawing the blueprints, but still
- 7:55missing that human artistic soul. That is the
- 7:57perfect way to look at it. Which naturally makes
- 7:59me wonder about its performance in a purely 2D
- 8:02space. Right. If it has perfect blueprints, but
- 8:05no interior designer in complex 3D environments,
- 8:08what happens when it tries to design a flat 2D
- 8:11website? How does it stack up against its direct
- 8:13predecessor, GPT -5 .5? So the community actually
- 8:16ran a brilliant direct comparison to test exactly
- 8:18that. They used a highly detailed spaceship design
- 8:21prompt. and the results exposed a major paradox.
- 8:25GPT 5 .6 Pro generated the image, but it ran
- 8:28for 87 minutes. Wait, 87 minutes for one single
- 8:31visual prompt? Yep, 87 minutes of continuous
- 8:34generation time. To put that in perspective,
- 8:37the older GPT 5 .5 model, running on extra high
- 8:40intelligence, completed the exact same prompt
- 8:43in 34 minutes and 42 seconds. I really have to
- 8:46point out the paradox there. An 87 -minute runtime
- 8:48isn't a feature you put on a marketing brochure.
- 8:51Definitely not. That represents a massive, almost
- 8:54unsustainable computing cost. If an AI takes
- 8:57an hour and a half to think through a visual
- 8:59prompt, the sheer volume of GPU cycles burning
- 9:02in the background is staggering. How do you scale
- 9:05an API that takes 87 minutes to answer one user?
- 9:08You don't. At least not yet. This massive resource
- 9:12burn is likely why it is hidden behind the A
- 9:14-B test rather than rolled out to everyone. That
- 9:16makes sense. But we have to look at what that
- 9:1887 minutes actually bought them. 5 .6 Pro definitely
- 9:22won on the micro -details, the specific lighting,
- 9:24the metallic shading on the spaceship, the intricate
- 9:27detail on the captive's chairs, and the exterior
- 9:29hull. Okay. It also produced far fewer visual
- 9:32glitches or warped pixels. But it wasn't a clean
- 9:35sweep. Right. GPT -5 .5 actually produced better
- 9:38interior rooms and far more compelling planets
- 9:40in the background. And again, the rival model,
- 9:43Fable 5, still beat both of OpenAI's models on
- 9:46the overall cohesiveness of the spaceship design.
- 9:50It really sounds like 5 .6 Pro is just an incremental
- 9:52update here, not the Fable 5 killer everyone
- 9:55was hoping for. For raw blank canvas created
- 9:58design, yes. It is purely incremental. Right.
- 10:01But the testers uncovered a completely different
- 10:03strength when they moved to design mimicry. They
- 10:06handed 5 .6 Pro a single reference image of an
- 10:10existing e -commerce landing page, and the model
- 10:13recreated it flawlessly. It nailed. the grid
- 10:16layout, the typography, the exact stylistic vibe
- 10:19of the original reference. I have to admit, I
- 10:21still wrestle with getting AI to match a specific
- 10:24design template. The prompt drift is real. Oh,
- 10:27it really is. You ask for a minimalist blue button,
- 10:29and three prompts later, the AI has decided your
- 10:31whole website should be neon purple. So seeing
- 10:34a model lock onto a visual template and hold
- 10:36it perfectly is genuinely impressive. It's huge
- 10:38for front -end developers. But does its success
- 10:40with the e -commerce page mean its true strength
- 10:42is just mimicry? This raises a critical question
- 10:44about the future utility of these models. Right
- 10:47now... Its absolute superpower in the 2D visual
- 10:51space is strict replication. OK. It performs
- 10:54exponentially better when you provide rigid guardrails
- 10:57and clear visual references. When you ask it
- 10:59to create from a pure blank page, it struggles
- 11:02to make cohesive stylistic choices. It needs
- 11:05you to define the aesthetic boundaries first.
- 11:08Better at strictly following the instructions
- 11:10than inventing a brilliant design from scratch.
- 11:13That is the reality of its current architecture,
- 11:15yeah. We will be right back after a quick word
- 11:17from our sponsor. Stick around. All right. And
- 11:20we are back. So we have established that this
- 11:23model needs strict instructions to design a standard
- 11:26website. But there is one highly specific visual
- 11:29format where it doesn't need to mimic, a format
- 11:32where it is genuinely shocking the developer
- 11:34community. SVG generation. This is the undisputed
- 11:38hidden superpower of the leaked checkpoint. Just
- 11:41to define that for a moment, SVG stands for scalable
- 11:43vector graphics. Simply put, it means scalable
- 11:46images drawn using mathematical formulas instead
- 11:48of individual pixels. Exactly. And because they
- 11:51are entirely mathematical formulas, generating
- 11:54complex lighting or shading is incredibly difficult.
- 11:56Right. When an AI draws a normal JPEG, it's basically
- 11:59just placing a dark pixel next to a light pixel
- 12:02based on training data. But with an SVG, the
- 12:04AI has to write pages of raw code to define gradients,
- 12:08light sources, and geometry. And the demo that
- 12:11broke the internet here was a BMW M4 CS. Yes.
- 12:155 .6 Pro rendered an SVG of this car, featuring
- 12:18metallic shading, correct lighting reflections,
- 12:22and flawless perspective. It looked astonishingly
- 12:25close to a photograph, purely driven by math.
- 12:27They ran a direct head -to -head comparison against
- 12:30Fable 5. They pushed Fable 5 across its low,
- 12:33medium, high, and extra -high thinking levels.
- 12:35And what happened? Fable 5 failed entirely. It
- 12:37could only produce flat, cartoonish vector styles.
- 12:40It simply couldn't do the high -level metallic
- 12:42math. That is a definitive victory. And it wasn't
- 12:44just a car. Another tester prompted it to generate
- 12:47a Windows 11 interface. This one was crazy. It
- 12:49recreated the full operating system UI in SVG
- 12:53format. The file explorer, the task bar, the
- 12:57calculator app. all mathematically drawn. It
- 13:00cleanly outclassed another specialized model
- 13:02called Mythos. It did. But, as always with this
- 13:06checkpoint, there was a strange downside. There
- 13:08always is. It hallucinates extra interface elements.
- 13:11During the Windows 11 generation, it added bizarre
- 13:14unnecessary pop -ups and lines of text that simply
- 13:18do not exist in the real operating system. It's
- 13:21like an overly eager intern who gives you the
- 13:23pristine, incredibly complex spreadsheet you
- 13:26asked for, but then decides to add 10 confusing
- 13:29pie charts that you didn't need just to prove
- 13:31they could. Yeah. Why does a model this mathematically
- 13:33advanced throw in fake pop ups and random text?
- 13:36Because the model is heavily optimizing for extreme
- 13:39detail. It equates visual density with quality.
- 13:41Oh, I see. It has all this incredible processing
- 13:43power and structural understanding, but it completely
- 13:47lacks editorial restraint. It doesn't know when
- 13:49a design is actually finished and should just
- 13:51be left alone. Incredible attention to detail,
- 13:53but severely lacking an editor's restraint. It
- 13:55just wants to keep painting until the canvas
- 13:57is entirely full. So bringing all these bizarre
- 13:59technical quirks, the massive 87 -minute generation
- 14:03times, and these undeniable SVG superpowers together,
- 14:07where does this leave us in the broader AI arms
- 14:09race? Well, if we look at the honest benchmark
- 14:13comparison, 5 .6 Pro absolutely dominates on
- 14:17SVGs. It dominates on vision replication. And
- 14:19it wins heavily on deep game logic and code stability.
- 14:23OK. But it still trails Fable 5, as well as Claude,
- 14:27another major competitor, on standard front -end
- 14:29web generation and overall aesthetic polish.
- 14:32We should also caveat that Opus, another heavyweight
- 14:34model in the industry, wasn't benchmarked in
- 14:36this specific leak. And then there are the rumors
- 14:38floating around the edges of the technical data.
- 14:41The unconfirmed market noise. Pricing is a huge
- 14:43topic of speculation. The rumor mill strongly
- 14:46suggests the cost will sit somewhere right between
- 14:48Fable 5 and Opus 4 .8, while magically matching
- 14:51the price of the older GPT 5 .5 model. But the
- 14:54code name confusion is where the data gets incredibly
- 14:56muddy. Testers are tracking an entire constellation
- 14:59of names. You've got Iris Alpha, Ember Alpha,
- 15:02Beacon Alpha, Kepler, and Kindle. And strangely,
- 15:06some testers are reporting that the checkpoint
- 15:08named Kindle Alpha actually performs worse than
- 15:10the one named Kepler. Yet Kindle is supposedly
- 15:14the finalized release candidate. I really have
- 15:17to push back on the logic of that specific rumor.
- 15:19If Kindle Alpha is verifiably performing worse
- 15:22in testing, Why on earth would a multi -billion
- 15:26dollar company make that the flagship release
- 15:28candidate? Doesn't make sense. It really highlights
- 15:31how messy and contradictory these secondhand
- 15:33leaks really are. We have to treat the code names
- 15:36with extreme skepticism. The code names are a
- 15:39distraction, honestly. The underlying behavioral
- 15:41shift is the only thing that actually matters
- 15:43here. So zooming out and looking at the landscape
- 15:45today, is this hidden checkpoint the fable five
- 15:48killer the industry has been waiting for? Not
- 15:51entirely. It is closing the technical gap at
- 15:53a terrifying speed, especially regarding structural
- 15:56logic and complex SDG math. But it is absolutely
- 16:00not taking the creative crown across the board.
- 16:02Closing the distance fast, but definitely not
- 16:04taking the crown just yet. Exactly. So what does
- 16:06this all mean? Let's synthesize everything we've
- 16:08unpacked today. The overarching theme is that
- 16:11we are witnessing a live, highly quiet evolution.
- 16:15of artificial intelligence happening right inside
- 16:18our daily tools. The sheer fact that OpenAI can
- 16:21run this massive A -B test on live accounts without
- 16:24a single announcement shows how fluid and continuous
- 16:27this technology has become. And the specific
- 16:30capabilities of GPT 5 .6 Pro prove that a major
- 16:33historical milestone has been reached. Complex
- 16:36logic, deep physics, structural object permanence,
- 16:40building a one -shot HTML game that maintains
- 16:42state tracking for 48 minutes, these are no theoretical
- 16:45challenges, these are now solved problems for
- 16:47AI. The structural foundation is built, which
- 16:50means the new ultimate frontier for artificial
- 16:52intelligence isn't just raw computational capability
- 16:56anymore, it is taste. It is editorial restraint.
- 16:59It is knowing how to make a digital environment
- 17:01feel distinctly human rather than just functionally
- 17:04correct. It is the classic difference between
- 17:06building a house and creating a home. The AI
- 17:10can build the house perfectly now, but making
- 17:12it feel lived in? That is the next great technological
- 17:16leap. I highly encourage you to check your own
- 17:18dashboard. Switch your model over to GPT 5 .5.
- 17:22Set your intelligence slider to high. See if
- 17:24you can spot the slower, significantly sharper
- 17:27responses of that ghost checkpoint for yourself.
- 17:30experience that untainted A -B test firsthand.
- 17:33It's definitely worth trying. I want to leave
- 17:34you with the lingering thought to chew on. We
- 17:36talked about that ghost in the machine. If an
- 17:38AI can now quietly build complete physics -based
- 17:413D worlds and functional simulation games in
- 17:44a single file, just by us asking, what happens
- 17:47to the value of human coding when that ghost
- 17:49finally develops real taste? That is the million
- 17:52dollar question. Until next time, keep diving
- 17:55deep.