Latest / AI Fire Daily / 🎙️ EP 303: Fable 5 & Mythos 5 Finally Restored & OpenAI’s GeneBench-Pro
Transcript
- 0:00Imagine reaching for your smartest tool right
- 0:02in the middle of a project, the one you rely
- 0:04on every single day to think alongside you. And
- 0:07suddenly, it is just gone. A total digital blackout
- 0:11that leaves you staring at a blank screen. Poof.
- 0:15Vanished. 19 days of pure silence from one of
- 0:18our most capable systems. Exactly. For nearly
- 0:20three weeks, that was our reality. But today,
- 0:23the lights are finally back on. Yeah. And the
- 0:25implications are absolutely massive. Welcome
- 0:28to today's Deep Guys. It is July 1st, 2026. We
- 0:33have a really fascinating stack of tech updates
- 0:35to sift through today. Right. Our central mission
- 0:38is exploring the messy reality of advanced AI.
- 0:41We are going to unpack the dramatic global return
- 0:44of Claude Fable 5. Finally. Yeah. Then we will
- 0:47look at the ballooning unsustainable costs of
- 0:49running these digital brains. After that, we
- 0:52will examine a flood of new consumer tools hitting
- 0:54the market. Which is a lot. It is. And finally,
- 0:57we will ask if AI can actually do complex biological
- 0:59science. It is a packed session. The landscape
- 1:02shifts so incredibly fast right now. Okay, let's
- 1:04unpack this. Let's start with the biggest news
- 1:06of the day. There is a massive tension. between
- 1:10making AI incredibly smart and keeping it safe.
- 1:13Right. Claude Fabel 5 and Mythos 5 Access is
- 1:16officially restored globally today. The U .S.
- 1:19government finally lifted those strict export
- 1:21controls. The entire system vanished back on
- 1:24June 12th for a very specific reason. Amazon
- 1:28researchers discovered a critical architectural
- 1:31vulnerability deep in the model's logic. They
- 1:35found a jailbreak. Which is tricking an AI into
- 1:37ignoring its own safety rules. Right. But this
- 1:40was not just getting it to say something slightly
- 1:42inappropriate. No, this specific jailbreak was
- 1:44highly dangerous. It let Fable 5 completely bypass
- 1:48its core safeguards. Wow. Yeah. It could actually
- 1:51write malicious zero -day exploit code. And the
- 1:53mechanics of that are what forced Anthropic to
- 1:56pull the plug. They could not just patch it on
- 1:57the fly for specific high -risk regions. Right.
- 2:00They could not verify user nationality in real
- 2:02time with enough certainty. They just shut access
- 2:05down for everyone globally. They had to act immediately.
- 2:08The potential fallout was simply too severe to
- 2:10risk. It is like recalling every car on Earth
- 2:12because someone bypassed a speed limiter. The
- 2:16scale of that decision is staggering when you
- 2:19really think about it. It really highlights how
- 2:21fragile our reliance on centralized models has
- 2:24become. But they finally engineered a fix. Anthropic
- 2:28worked directly with U .S. government security
- 2:30researchers to build a solution. They did not
- 2:33just retrain the main model, though. They built
- 2:35a separate system entirely. Exactly. They trained
- 2:38a new dedicated classifier that sits in front
- 2:41of the model. OK. It analyzes the incoming prompt
- 2:44before Fable 5 even starts. sees it. It blocks
- 2:46that specific jailbreak technique in over 99
- 2:49% of cases. Wow. Yeah. Government researchers
- 2:52verified the new guardrails are extraordinarily
- 2:54strong. I have to make a vulnerable admission
- 2:56here. I still wrestle with prompt drift myself.
- 3:00Oh, sure. Getting these models to do exactly
- 3:03what you want safely is incredibly hard. You
- 3:05give it a rigid set of rules, but it slowly loses
- 3:08focus. It is a constant tug of war for developers
- 3:11right now. You want maximum capability, but you
- 3:14absolutely need those rigid guardrails in place.
- 3:17And those guardrails are creating a deeply frustrating
- 3:19experience for heavy users. Yeah. If you are
- 3:22a pain user relying on Fable 5... There's a massive
- 3:26catch. Yeah, you get Fable 5 globally on web
- 3:29and mobile right now. But there are very tight
- 3:32new usage limits attached to your account. You
- 3:35get a strict 50 % quota through July 7. Because
- 3:38Fable 5 is running massive parallel inferences
- 3:40in the background. It draws down your token usage
- 3:42a lot faster than Opus 4 .8 ever did. Right.
- 3:45After July 7, or if you hit the cap early, you
- 3:47need credits. You really have to plan your workflows
- 3:49around these strict new limits. You cannot just
- 3:52spam complex queries to brainstorm anymore. Plus,
- 3:55the new safety classifier creates some serious
- 3:57workflow friction. If your prompt even brushes
- 4:00up against a safeguard, it gets automatically
- 4:02rerouted. You get bumped straight down to Opus
- 4:044 .8 without warning. And the classifier is aggressively
- 4:08tuned right now to prevent another disaster.
- 4:11A small fraction of routine, completely benign
- 4:14coding tasks might trigger a false positive.
- 4:17That kind of unpredictability is a nightmare
- 4:20for enterprise software development. I am wondering
- 4:23how heavy users are actually going to handle
- 4:25these tight new quotas. They will have to be
- 4:27highly strategic. You have to cash your results
- 4:30and stop running redundant queries. So heavy
- 4:32users hit limits quickly or get silently bumped
- 4:34down to older models. Precisely. It is a brand
- 4:37new era of strict AI rationing. Which perfectly
- 4:40bridges into our next topic. This massive effort
- 4:43to secure and run these models brings up a fundamental
- 4:47issue. The infrastructure costs for these companies
- 4:50are absolutely skyrocketing. It is becoming the
- 4:52defining bottleneck of the entire tech industry.
- 4:55Business leaders are getting noticeably frustrated
- 4:57by the current economic reality. Palantir CEO
- 5:01Alex Karp just came out swinging against the
- 5:04system. He is openly criticizing open AI and
- 5:07Anthropix's reliance on the token model. The
- 5:09token model. That is paying for AI by the chunk
- 5:12of words it processes. Right. Karp makes a very
- 5:16blunt, practical point about enterprise deployment.
- 5:19Okay. Companies are completely exhausted by these
- 5:22endlessly rising, unpredictable AI costs. They
- 5:26feel like they're hemorrhaging cash on compute
- 5:27and want their control back. I totally hear that.
- 5:30It is like stacking Lego blocks of data. The
- 5:33tower just keeps growing and the monthly bill
- 5:36never actually ends. Right. When an agent runs
- 5:38thousands of hidden queries. The token model
- 5:41breaks down. Exactly. You ask one simple question,
- 5:44but the agent runs 50 API calls. Your token usage
- 5:48explodes before you even get a single usable
- 5:50answer. Are enterprise companies even seeing
- 5:52actual ROI on these massive compute bills yet?
- 5:56That is the billion dollar question driving the
- 5:58entire market right now. The financial sector
- 6:00is reacting aggressively to this exact tension.
- 6:04Metashares just jumped nearly 9 % on a very telling
- 6:07rumor. Reports suggest they are actively building.
- 6:10building a massive new cloud business. They want
- 6:12to sell their excess AI compute to other desperate
- 6:14companies. It makes perfect sense from a capital
- 6:16expenditure standpoint. Meta bought hundreds
- 6:20of thousands of GPUs to train their own open
- 6:22source models. Right. Now they are sitting on
- 6:24a mountain of raw compute power. Turning that
- 6:28massive sunk infrastructure cost into potential
- 6:31new revenue is brilliant. It shows how desperate
- 6:34the broader market is for any cheaper infrastructure
- 6:37alternatives. Look at the recent fundraising
- 6:40numbers in the alternative compute space. A company
- 6:43called Together AI just raised $800 million in
- 6:46fresh capital. Wow. Yeah. Their valuation just
- 6:49skyrocketed to $8 .3 billion almost overnight.
- 6:53And they specifically focus on providing cheaper,
- 6:55more efficient AI infrastructure. Yeah, they
- 6:58optimize the hardware layer and provide open
- 7:00source model hosting. They are meeting this massive,
- 7:03desperate market demand for predictable operating
- 7:06expenses. When you run an open source model on
- 7:08your own rented servers, costs stabilize. You
- 7:12are not paying a premium tax on every single
- 7:14word generated. Exactly. Is open source the only
- 7:16way companies can escape this token billing trap?
- 7:19It is definitely the most viable long term path
- 7:22forward right now. Owning the model means you
- 7:25have predictable manageable monthly costs. You
- 7:29finally stop paying rent on every single query
- 7:32your company runs. Essentially businesses want
- 7:35their own keys to the car not an endless taxi
- 7:38ride. Exactly. They need to drive their own economic
- 7:41destiny. Despite these massive crippling infrastructure
- 7:44costs, the pace is not slowing down at all. Tech
- 7:48companies are churning out new tools in an absolute
- 7:51blistering pace. Oh, yeah. They have to release
- 7:54constant updates just to justify the capital
- 7:56spend. It is an absolute flood of new capabilities
- 7:59hitting the market this week. We are seeing a
- 8:01massive transition toward fully autonomous digital
- 8:03workers. Here is where it gets really interesting.
- 8:06Looking at all these new tools, it feels like
- 8:08we are about to drown. I cannot imagine. manually
- 8:12managing a dozen different bots just to get through
- 8:15work. That is the secret, though. They are designing
- 8:18these tools to eventually orchestrate themselves
- 8:20without you. Let's look at the underlying thread
- 8:22connecting all these new software releases. It
- 8:25is all about moving from text generation to cross
- 8:28-modal synthesis. Right. Notebook LM just dropped
- 8:31video overviews, and the mechanics are fascinating.
- 8:33They take complex source documents and synthesize
- 8:36them into 60 -second clips. Wow. It generates
- 8:38relevant images and voiceovers perfectly synced
- 8:41to the core concepts. Doom scrolling finally
- 8:44becomes somewhat educational. Making doom scrolling
- 8:47educational is a brilliant psychological pivot
- 8:50for consumer tech. It takes our worst habits
- 8:53and wraps them in high -density knowledge transfer.
- 8:56Yeah. And Google is aggressively pushing this
- 8:58cross modal capability. They launched Gemini
- 9:013 .5 Live Translate for seamless, real -time
- 9:05audio translation. They teased deep Gemini integration
- 9:08coming natively in Android 17. And they introduced
- 9:12the new Google Home speaker built entirely around
- 9:15Gemini's architecture. They are not just launching
- 9:17apps anymore. They are baking the AI directly
- 9:20into the operating system layer itself. Exactly.
- 9:22And the models driving these features are becoming
- 9:25incredibly powerful. Claude Sonnet 5 is officially
- 9:28out and it is a major leap forward. Anthropic
- 9:31explicitly calls it their most agentic Sonnet
- 9:35model yet. Agentic software. That means software
- 9:38that acts on its own to achieve your goals. Right.
- 9:41It operates semi -autonomously to solve complex
- 9:44coding and professional workflows. Then there
- 9:46is Gemini OmniFlash, which is pushing the boundaries
- 9:50of video. It supports high -quality, real -time
- 9:53video generation, but it also allows conversational
- 9:56editing from mixed text, image, and video inputs.
- 9:59You can literally just talk to your video editor
- 10:01to make complex cuts. We are also seeing a major
- 10:05structural shift in how apps actually communicate.
- 10:07Look at a new tool called Acti. It is an agentic
- 10:10keyboard designed specifically for mobile commands
- 10:12and search. It bypasses the traditional app interface
- 10:15completely. It understands your intent and fetches
- 10:17the right data directly into your text box. And
- 10:20then there is Humalike, which is fascinating
- 10:22from a psychological perspective. They are building
- 10:24behavioral infrastructure for human -like AI
- 10:27agents. They provide the nuanced social skills
- 10:30and proactiveness that raw models currently lack.
- 10:32They are essentially selling digital empathy
- 10:34as an API service. Two sec silence. But the most
- 10:39technically impressive tool this week might be
- 10:41tab stack browser automation. TabStack is a total
- 10:45game changer for agentic workflows. Traditionally,
- 10:49automating the web required hosting clunky, memory
- 10:52-heavy, headless browsers. Yeah, which is a pain.
- 10:56It is. TabStack completely rips out the visual
- 10:58rendering engine. Yeah. It just automates the
- 11:00web within your app and streams the DOM state
- 11:02back. You do not even need to host a browser
- 11:05instance anymore. The result just streams back
- 11:07seamlessly in one incredibly efficient API call.
- 11:11That kind of efficiency is exactly what - drives
- 11:13the cost down. But with all this rapid, chaotic
- 11:16growth, the industry is swarming with rumors.
- 11:19Oh, yeah. There is an unverified list circulating
- 11:21that nine major U .S. tech companies are pivoting.
- 11:24The rumor claims they are actively exploring
- 11:26Chinese alternatives for their infrastructure.
- 11:29Whether it is true or just market posturing,
- 11:31it is incredibly telling. We are just reporting
- 11:33on the rumor here, of course, but it highlights
- 11:35just how desperate companies are for cheaper
- 11:37compute alternatives globally. Absolutely. Another
- 11:39persistent rumor involves the recent rollback
- 11:41of Claude Code features. There were accusations
- 11:44about hidden China proxy and time zone signals
- 11:48inside system prompts. The mechanics of that
- 11:51are interesting. People thought the model was
- 11:54secretly routing data through unauthorized geographic
- 11:57regions. The good news is, anthropic staff immediately
- 12:00addressed the controversy publicly. They confirmed
- 12:03that specific prompt writing behavior is already
- 12:06being actively rolled back. And finally, there
- 12:09is a fascinating rumor in the hardware space.
- 12:12SpaceX supposedly showed investors a prototype
- 12:14for a handset -like AI device. A physical AI
- 12:18phone that is reportedly much slimmer than a
- 12:20standard iPhone. Though we should note, Elon
- 12:22Musk quickly stated the rumor is false. Right.
- 12:24But even if it is early, it shows everyone is
- 12:26rushing toward ambient AI hardware. With all
- 12:31this happening, how are everyday people supposed
- 12:33to juggle these tools? The tools will eventually
- 12:35start talking directly to each other via APIs.
- 12:38You will not juggle them manually or open 12
- 12:40different apps. You will just give high -level
- 12:42goals, and the agents will orchestrate the messy
- 12:44details. We're moving from typing manual commands
- 12:47to just supervising AI assistants doing the work.
- 12:50That is the exact future we are rapidly speeding
- 12:52toward. Sponsor! So what does this all mean?
- 12:56We have all these agentic tools and incredibly
- 12:58fast cross -modal models. Yeah. But a fundamental
- 13:01lingering question remains about their actual
- 13:05underlying intelligence. Can they actually do
- 13:08the complex, messy, intuitive work of a human
- 13:11scientist? OpenAI just released a massive new
- 13:14benchmark to test exactly that limit. It is called
- 13:16GenoBench Pro, and it is a completely new standard
- 13:20for scientific AI. It focuses heavily on computational
- 13:23biology and advanced genomics. But it goes far
- 13:26beyond standard multiple choice logic tests.
- 13:29Right. It actively evaluates something researchers
- 13:31call research taste. Research taste. That is
- 13:34the intuition to pick the right path through
- 13:36messy data. It is not just about memorizing static
- 13:39biological facts from a textbook. It is about
- 13:41making complex judgment calls when the data contradicts
- 13:44itself. Real world science is incredibly messy,
- 13:47full of ambiguity and constant false signals.
- 13:49The benchmark features 129 incredibly complex,
- 13:53multi -layered problems. They span across genomics,
- 13:57translational medicine, and deep quantitative
- 13:59biology. The models are not given a clear path
- 14:02to the answer. They must independently figure
- 14:05out how to structure their own analytical approach.
- 14:07And the mechanics of how an AI simulates that
- 14:10intuition is fascinating. It relies on probabilistic
- 14:13weighting of massive data vectors. It has to
- 14:16look at a messy genome sequence. and make a choice.
- 14:19It has to actively distinguish between actual
- 14:21biological phenomena and just statistical noise.
- 14:25That is notoriously difficult because sequencing
- 14:27machines inherently produce random technical
- 14:29errors. A human researcher knows to ignore an
- 14:32artifact because the machine runs hot. How does
- 14:35an AI learn that implicit, undocumented contextual
- 14:38knowledge? Exactly. OpenAI had 82 of these questions
- 14:42rigorously reviewed by external domain experts.
- 14:45They needed to verify they actually represent
- 14:47real work. laboratory conditions the human reviewers
- 14:50came back with a striking estimation of the difficulty
- 14:52they concluded that a single problem would take
- 14:54a trained expert days to solve yeah roughly 20
- 14:57to 40 hours of dedicated cognitive labor per
- 15:00question to sex silence that is a massive amount
- 15:04of sustained human effort How did the models
- 15:07actually perform under that kind of intense pressure?
- 15:10OpenAI's new GPT -5 .6 Sol model currently leads
- 15:13the pack. In its advanced Pro mode, it hit a
- 15:1631 .5 % pass rate. Claude Opus 4 .8 came in second,
- 15:21hitting 16 .0%. And Gemini 3 .5 Flash trailed
- 15:25behind at just 8 .1%. A 31 .5 % pass rate is
- 15:29certainly an interesting milestone. But there
- 15:31is still a massive glaring gap between that and
- 15:34actual science. Oh, for sure. Whoa. Imagine scaling
- 15:43that kind of analytical power to a billion queries.
- 15:46Even at a 31 % success rate, the sheer volume
- 15:48of discovery is staggering. It really is. For
- 15:53transparency, OpenAI is open sourcing a subset
- 15:58of these complex benchmark questions. Independent
- 16:01verification is absolutely crucial when evaluating
- 16:04these massive proprietary claims. Beat. Is 31
- 16:09.5 % actually impressive for a machine given
- 16:12that 40 -hour human time frame? Absolutely. It
- 16:15fundamentally alters the math of scientific research.
- 16:18A human takes a whole week to either fail or
- 16:21succeed at one problem. Yeah. And AI runs the
- 16:23simulation in mere minutes. The iterations happen
- 16:26so fast, the successful discoveries compound
- 16:28incredibly quickly over time. Right. They fail
- 16:31often, but when they succeed, they save entire
- 16:33work weeks. Exactly. It completely changes the
- 16:36underlying pace of biological discovery. We have
- 16:38covered a massive amount of complex ground today.
- 16:40Let's pull back and look at the broader trajectory
- 16:42of where this is heading. all these individual
- 16:44updates, the bigger picture is truly fascinating.
- 16:47We are watching AI evolve in real time right
- 16:49in front of us. It started as a deeply fragile,
- 16:52easily manipulated system, one that required
- 16:55government bans and massive, incredibly expensive
- 16:58server farms just to operate safely. And now
- 17:01it is slowly transforming into a highly capable,
- 17:05semi -autonomous entity. It is actually beginning
- 17:08to develop the probabilistic intuition to pass
- 17:11graduate level genetics exams. It is a very messy,
- 17:15very chaotic transition period for the industry.
- 17:18The compute costs are unsustainably high and
- 17:21the agentic tools are proliferating wildly. Yeah.
- 17:24But the underlying capability of these models
- 17:27is completely undeniable. We are actively building
- 17:30our future digital colleagues right now. Which
- 17:32leaves us with a final, slightly unsettling thought
- 17:35to mull over. If an AI can be reliably benchmarked
- 17:39on its research taste. If it can navigate the
- 17:42noise to answer incredibly complex biological
- 17:44questions, what happens when it develops a taste
- 17:46for questions we have not even thought to ask
- 17:48yet? That is a thought that will definitely keep
- 17:50you up at night. Thank you for joining us for
- 17:52this deep dive session. Remember to check your
- 17:54usage credits before your agents drain them.
- 17:56Good luck with that. Stay curious out there.
- 17:58And be sure to subscribe for our next session.
- 18:00Out TOO music.