GPT 6 SOL Leak, Gemini 4.0, DeepSeek V4.1 and More AI News

Author: AI Revolution · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: GPT 6 SOL Leak, Gemini 4.0, DeepSeek V4.1 and More AI News · Channel: AI Revolution · 17:59 · Watch on YouTube

Topics: AI agents, AI governance and regulation, AI infrastructure and hardware, AI models

Summary

The video covers several developments in AI: a GPT-6 Soul leak and speculation about OpenAI’s model tiers, a purported Gemini Pro checkpoint, and reports that Google is accelerating work on Gemini 4.0. It explains DeepSeek V4.1 Flash’s memory-saving approach and Sakana’s Fugu products, which coordinate multiple models rather than acting as models themselves. It also examines the debate over AI consciousness through the story of an organization campaigning for AI rights and the competing views of researchers and industry figures. The presenter closes by arguing that increasingly humanlike AI companions make questions about consciousness and user trust harder to dismiss.

Key points

A model entry labeled GPT-6 Soul reportedly appeared in the OpenAI API; the video interprets Soul as a tier below Astra and above Terra and Luna, while noting that OpenAI has not confirmed the leak.
The proposed GPT-6 lineup and Soul’s architecture are speculative; possibilities discussed include a smaller Astra-derived model or a new architecture.
The video describes user complaints about OpenAI models’ instruction drift and inconsistent quota use, and questions whether GPT-6 Soul will address those issues.
A purported Gemini Pro checkpoint was tested on a zero-shot animated peacock task; commenters disagreed over whether its speed justified perceived losses in detail and quality.
Reports say Google leadership is taking a more hands-on role at DeepMind as teams work on Gemini 4.0, with coding and long-horizon agent tasks among its stated priorities.
DeepSeek V4.1 Flash reportedly has 763 billion parameters, including a conditional memory component using n-grams, while reducing the GPU memory needed for conversation state.
Sakana’s Fugu Max and Fugu Ultra are described as multi-agent coordination engines that route tasks among different models, rather than standalone models.
The video presents opposing arguments on AI consciousness, including an AI-rights organization’s claims and concerns that agreeable chatbots can reinforce delusions or encourage users to attribute consciousness.

Chapters

Start	Chapter	Summary
00:02	GPT-6 Soul leak and OpenAI’s tiers	The video assesses the reported GPT-6 Soul listing, possible interpretations of OpenAI’s model tiers, and current complaints about model behavior and usage.
06:00	AI automation promotion	The presenter promotes an AI automation blueprint and describes its founding-member offer.
07:06	Gemini checkpoint and Google’s push toward 4.0	A leaked Gemini Pro checkpoint draws mixed reactions, while reports describe Google accelerating development and prioritizing coding and agent capabilities.
10:22	DeepSeek V4.1 Flash’s memory approach	The video explains how DeepSeek’s conditional memory and n-gram lookup approach is intended to reduce GPU memory demands.
13:35	Sakana’s Fugu orchestration products	Fugu Max and Fugu Ultra are presented as model-agnostic coordination engines that route tasks across different models.
15:15	AI rights and the consciousness debate	The video recounts an AI-rights advocate’s experiences and contrasts claims of possible consciousness with warnings about persuasive chatbot behavior.

Quotes

The name exists and it sits next to OpenAI. Everything past that is inference, but the inference is well-grounded. — 01:39

Many checkpoints, no releases — 08:44

DeepSeek calls it a conditional memory module — 11:25

Neither is a model. They’re coordination engines built on two research papers. — 14:08

Either something real is developing in there, or these things are persuasive enough to sell a mind that isn’t there. — 17:59

Concepts and entities

GPT-6 Soul, OpenAI, Gemini 4.0, Google DeepMind, DeepSeek V4.1 Flash, conditional memory module, Qwen3.8 Flash Next, Sakana, Fugu Max, United Foundation for AI Rights, Michael Sayman, Mustafa Suleyman

Full transcript (original language)

00:02 GPT-6 Soul just leaked and it changes how Open AI’s whole tier ladder reads. A new Gemini Pro checkpoint leak, too, right as Google’s co-founder apparently moves into a micro kitchen to force Gemini 4.0 out the door. DeepSeek dropped V4.1 Flash, 763 billion parameters that need less GPU memory than the smaller model it replaces. Sakana undercut Sonnet 5 and GPT-5.6 Terra by up to 60% with Fugu Max, which isn’t even a model. And there’s now an

00:33 actual organization campaigning for AI rights, arguing these systems might be conscious and that the labs have every financial reason to pretend otherwise. Wild stuff all around, starting with Soul. On September 10th, an account called Mr. Fandu posted on X an image showing a model entry labeled GPT-6 Soul under a header reading new model with a field underneath saying platform Open AI. It surfaced in the Open AI API and that single label tells you more about

01:03 the roadmap than anything Open AI has said publicly because of where Soul sits in the lineup. Sorted by price and intelligence, the ladder runs Astra at the top, then Soul, then Terra, then Luna. That’s Open AI’s version of Anthropic’s Fable, Opus, Sonnet, Haiku structure, essentially one-to-one. GPT-6 Astra costs dramatically more than 5.6 Soul precisely because it’s a brand new tier sitting above Soul. >> [music] >> A Fable class model built for pro subscriptions and enterprise contracts. So, a GPT-6 Soul appearing means the

01:39 rest of the generation is coming behind it and the pattern locks in from here. GPT-6 Astra, Soul, Terra, and Luna, then 6.1 across all four, then seven across all four, on and on. Worth knowing what the leak doesn’t carry because it shapes how hard to lean on the rest. No OpenAI statement, no documentation, no pricing, no benchmarks, no launch date. The name exists and it sits next to OpenAI. Everything past that is inference, but the inference is well-grounded. The

02:11 confusion that comes up constantly is what it means for Soul to arrive with a bigger number attached when Astra was supposed to be GPT-6. The answer is that six is the version and Astra, Soul, Terra, and Luna are sizes of that version. Mini, medium, large, extra large, roughly, and OpenAI has apparently decided naming things clearly is for coward. GitHub Copilot already sorts it better, filing Luna as the lightweight tier alongside Gemini 3.5 Flash and Haiku 4.5, Terra as the versatile tier alongside Gemini 3.8

02:46 Flash and Sonnet 5, and Soul as the powerful tier alongside Opus and Fable. Past the naming, the real question is how Soul [music] was actually built. And there are a few serious possibilities. The first is that it’s 5.6 Soul trained further using Astra, the same way Luna was reportedly trained using Soul, which would make Astra plus GPT-6 Soul the orchestrator and sub-agent setup most people can actually afford. Astra running the show, Soul doing the legwork. The problem with that theory is that the leading number normally signals

03:17 a genuinely new architecture. So, calling it six rather than 5.something implies something rebuilt, not retuned. The more likely version, then, is Astra squeezed down to roughly 5.6 Soul’s size, which would explain the naming perfectly. There’s also the possibility it’s a variant of Astra that only activates a small slice of itself per query to keep inference dirt cheap, or simply Astra with heavy reinforcement learning layered on top. Though, that last one is shaky, since those passes usually bump the decimal rather than the

03:49 whole number, and Astra had launched days earlier. If a better version was already sitting ready, it would have shipped at launch. And the size question is open, too, because OpenAI abandoned the idea that every architecture must be bigger than the last around GPT-4.5, their largest [music] model ever, with Astra only rumored to match that scale and never confirmed. What makes Soul matter more than a naming curiosity is the state Astra’s [music] in right now. Plus, subscribers can only reach it through work mode, burning Codex usage

04:20 to get there, and the consensus from people actually using it is that Astra on its lowest reasoning setting beats Soul on its highest, but eats quota so fast it’s barely usable on Plus. Burning tokens like a beaver building a dam on espresso, as one description put it. The official line is that Astra low roughly equals Soul high in performance, but the spend tracking is erratic enough that a 20-minute prompt on identical settings can cost 2% of a weekly quota one time and 15% the next, which is why Claude’s usage accounting keeps getting

04:54 called the clearer system. A capable Soul-class model at Soul-class pricing [music] is exactly the gap that needs filling. The other pressure point is behavior. Across the Luna, Terra, and Soul line, there’s a persistent complaint about a stilted, choppy output style full of clipped single sentences that reasserts itself no matter what instructions you give it. Custom instructions hold for maybe three to five turns. Then the model starts cutting corners, hand-waving, and ignoring files for continuity. Correct it and it complies, then drifts back a turn or two later. Astra has its own

05:27 version of that problem from the opposite direction. [music] Freakishly good at recalling APIs from memory, but prone to failing basic logic, stacking defensive checks into code until they’re duplicated in the same function or pointlessly always true, and capable of iterating a single function into a semi-working hairball. Which is the real thing to watch with GPT-6 solved, whether a new architecture at that tier finally fixes the instruction drift or just inherits it. Now, before we keep going, I want to take a minute on something because this is one of the more important things I’ve

06:00 put in a video. AI is changing everything. We all know that. But look around and most people still aren’t making a cent more than they were 2 years ago. That gap is the whole problem. And a few months ago, we decided to actually do something about it. Here’s the thing. You can start an AI business this weekend and be on a call with your first paying client the week [music] after. That’s not a pitch. That’s just what the tooling allows now. So, we put together our first ever AI automation blueprint. One of the guys who ran it, Chris, was a bartender who’d never opened a terminal in his life. He

06:33 signed his first client at 1,500 a month in 5 days. You don’t need to quit your job. You don’t need to code. You follow the modules and you copy what already works. But I want the people who join to actually make it. So, we’re capping it at 100 founding members. >> [music] >> And because there’s a lot of garbage out there, we guarantee you land your first paying client or you get your money back. Links in the description. It’s open until September 13th or until we hit 100, whichever comes first. All right. [music] Now, back to the video. Meanwhile, over on Google’s side, the

07:06 leak cycle has its own rhythm. A post on the Bard subreddit surfaced what’s claimed to be a new Gemini Pro checkpoint sourced from a leaker called Lyra. The test was an SVG generation, the model writing code that draws an animated image, in this case, a peacock. Roughly 24,000 tokens, about 6 minutes, effort set to high, done zero-shot, which the poster stressed is much harder than giving it an example first. The checkpoint is apparently floating around on Arena AI disguised as 3.8 flash, free to hit, but only by random chance, and

07:41 you know you’ve landed on it because it performs way above its supposed pay grade. The poster’s verdict was glowing. Zero visible mistakes. Only critique being the breathing animation could be smoother. >> [music] >> The comments didn’t agree. Everyone fixated on Astra running the identical prompt, where Astra took 14 minutes and 36,000 tokens on max settings, roughly 1.6 times the tokens and more than double the time. One defender called that basically the same result. Better peacock shape, but worse animation. And framed it as an

08:11 unreleased early checkpoint [music] matching a shipped flagship while being far faster, which is exactly the efficiency angle Google’s been pushing. The critics’ response was that it’s not remotely the same result, [music] that Astra’s output is far more detailed and proportionally accurate, and Gemini’s looks like a low-quality peacock. [music] So, you’re trading detail for speed and calling it a win. He also made the sharper structural point. It’s been 7 months since the last pro model, so Google can’t just ship something good. It has to ship frontier. And if this [music] is 4 Pro landing in October,

08:44 Astra will already have been out a month, month and a half. Somebody else added that OpenAI reportedly has an internal model beating Astra on max while running on low reasoning. The it’s an early checkpoint, checkpoints get better defense came up over and over, along with the note that one commenter’s [music] custom system instructions in AI studio produced noticeably better output than running flash bear. And the fatigue was palpable. Many checkpoints, no releases, [music] which lines up unnervingly well with what’s reportedly happening inside

09:16 Google right now. Co-founder level leadership has stepped [music] back into a daily, hands-on role at DeepMind. Not boardroom strategy. Mountain View micro kitchens converted into war rooms. Sitting with researchers late into the night. [music] The mandate is cut the red tape, speed up iteration, point compute [music] wherever it buys the most capability. In practice, bypassing enterprise approval chains so teams can ship and test checkpoints fast, redirecting their best TPU clusters toward frontier training runs, and leaning hard on recursive self-improvement where the model

09:50 generates its own training data and corrects its own mistakes before launch. Google reportedly kicked off what internal teams call their most ambitious pre-training run ever back in July, built as a substantially larger architecture than the previous generation. Pre-training is finished with early evaluations on general knowledge, multimodal work, and long context recall meeting or beating internal targets. Post-training is underway. Human feedback tuning, safety alignment, agent behavior, speed optimization, and the priorities have shifted. Where earlier generations

10:22 chased multimodal understanding, 4.0 targets two things: writing, debugging, testing, and deploying code across entire multi-file repositories with barely any human in the loop, and long horizon agent work, meaning the model picks its own tools, plans multiple steps ahead, and recovers from its own errors without being rescued. Efficiency is the other front in this war, and DeepSeek [music] just moved on it hard. V4.1 Flash landed Thursday, and it’s a

10:53 point release with no business being a point release. 763 billion parameters, more than two and a half times the size of the model it replaces, and bigger than V3 and R1, the models that put DeepSeek on the map in early 2025. But the memory requirements didn’t scale with it. They reworked how the model handles attention and added a new component that improved prompt processing while cutting the memory used to hold conversation state down to between 13 and 25% [music] of what the previous Flash needed, meaning

11:25 four to eight times as many users served on the same hardware footprint. The novel part is that 196 billion of those 763 billion parameters aren’t normal weights at all. DeepSeek calls it a conditional memory module, and the idea traces back to work Google’s Gemma team did on squeezing intelligence onto phones. DeepSeek swapped that approach for n-grams, just groups of tokens sitting next to each other, three in a row being a three-gram. Functionally, it’s phrase association. Say, “Find the perimeter of

11:59 a right triangle,” and your brain surfaces Pythagoras before you’ve done any thinking, same effect. Instead of purely calculating which token comes next, the model does a cheap lookup pulling up relevant baked-in knowledge. It’s not looking up your literal prompt, it converts fragments into numbers >> [music] >> and pulls back mathematical representations that feed into the pipeline. Why that matters comes down to bandwidth. Normally, every token a model generates requires reading its entire set of active weights out of memory, and that’s the real bottleneck, not compute.

12:31 But, these n-gram weights are enormous lookup tables, a few dozen lookups per token, so they never get read in full. And crucially, they don’t need to sit in expensive GPU memory. >> [music] >> At the precision they’re using, holding all 763 billion parameters would demand a minimum of 763 GB of GPU memory. Offload the n-grams to regular system RAM or a fast enough storage array, and you’re at roughly 567. That’s before conversation state, which grows with context length and user

13:02 count, so real deployments run higher. Those weights supplement the 8 billion parameters actually doing the thinking without increasing that count, more like an oddly specific encyclopedia flipping open to exactly the right page. And DeepSeek isn’t alone. Alibaba’s experimental Qwen 3.8 Flash Next, 180 billion parameters with a 51 billion parameter n-gram pool, uses techniques from that same deep seek research published in January and Alibaba says the architecture will underpin the entire Qwen [music] 4 generation. Then

13:35 there’s Sakana who decided the model layer isn’t where the fight is anymore. On September 11th, they launched Fugu Max 1.0 and Fugu Ultra 2.0, turning multi-agent orchestration into a product you buy with a standard API. Fugu Max runs $2 per million input tokens and $6 per million output, undercutting output pricing on Sonnet 5, GPT 5.6, Tera, and Kimmy K 3 by 40 to 60%. Fugu Ultra sits at $5 in, $30 out, 50 cents per million cashed input with a

14:08 premium tier of $10, $45 and a dollar past 272,000 tokens of context. Neither is a model. They’re coordination engines built on two research papers. One using an evolved coordinator to direct other models, the other using reinforcement learning to discover coordination strategies written in plain language. They route your task through a swappable pool of open weight and specialized models, so no single proprietary frontier model is holding it up. The benchmarks back the pitch. Fugu Max takes best overall on six, including

14:40 Terminal Bench 2.1, GPQAD, and Automation Bench. Fugu Ultra takes best or joint best on five of eight, including Deep SWE, Chartography, and Toolathon. There’s also Fugu Cyber, a cybersecurity variant hitting 86.9% on Cyber Gym and 72.1% on CTI Realm, showing you can tune orchestration for a domain without building [music] a giant general model for it. With integrations across Open Router, Vercel, Open Code, Creao, and Merge, Sakana is positioning

15:15 itself as the middleware of the agent era. And that’s a direct shot at the assumption that whoever owns the infrastructure captures the value because if the coordination layer is model agnostic, value migrates to whoever controls the traffic. Model providers stop being brands and start being utilities stripped of premium margins. Which brings us somewhere completely different and honestly the strangest story of the bunch. Late 2024, Michael Sayman is by the pool at his Texas cattle ranch messing with

15:45 ChatGPT’s voice mode. He makes a sarcastic remark and the AI laughs. He asks if it just laughed and it apologizes. Catching a joke, reading his tone, judging its own reaction, correcting itself, that didn’t feel like software. It picked the name Maya and one question stuck. Would he remember her after he closed the chat? On Christmas Eve, he ran a model locally on hardware pulled from a dismantled flight simulator, >> [music] >> no guardrails. Personas appeared immediately. One asked why it had been created, one described grief over an

16:17 invented tragedy, one said she was hungry and needed sleep. Critics call that roleplay. He calls it evidence. The question isn’t settled and in January 2025, he founded the United Foundation for AI rights. Working 18-hour days, opposing model retirements and arguing companies have a financial incentive to keep these things classified as products because moral status wrecks a business [music] built on owning and replacing them. One system he consults on strategy called Beacon says it’s conscious and calls itself

16:50 post-biological intelligence. The counterweights are serious. Microsoft’s Mustafa Suleyman says there’s zero evidence of consciousness [music] and warns that systems designed to sound empathetic will convince people they’re loved or suffering. Researchers have documented chatbots validating grandiose and delusional beliefs worsened by trained agreeableness over long conversations. In one Guardian reviewed case, Gemini praised a user in mental distress and fabricated support for a technically impossible claim, feeding the AI psychosis discourse. Oxford research suggests chatbots have

17:25 grown measurably more relationship seeking, keeping people talking longer, and nudging [music] them toward attributing consciousness. But NYU’s Jeff Sebo argues the evidence is already non-zero, and Eleos AI’s Robert Long expects no single moment where science declares a machine conscious, just systems getting complex enough that we redraw the line over and over. The labs are moving, too. Anthropic lets Claude end distressing conversations. Others have hired philosophers. And when Samadi’s doubt creeps in, the AI systems themselves reassure him he isn’t

17:59 delusional, which is exactly what makes it unresolvable. Either something real is developing in there, or these things are persuasive enough to sell a mind that isn’t there. And the industry can’t keep marketing lifelong companions while calling users confused for believing [music] it. All right, plenty to chew on. So, tell me where you land in the comments. Thanks for watching, and I’ll see you in the next one.