Latest / AI Fire Daily / 🎙️ EP 295: OpenAI Becomes an Ad Org at Cannes & Claude Sonnet 5 "Fennec" Rumors Heat Up
Transcript
- 0:00Imagine waking up tomorrow morning. beat. The
- 0:03sun is just rising outside your window. And your
- 0:06AI didn't just answer some questions overnight.
- 0:08It actually ran your entire business while you
- 0:10slept. It negotiated contracts, right? And optimized
- 0:13your campaigns. Exactly. Welcome to today's deep
- 0:17dive. We are looking at a monumental industry
- 0:19shift today. We've curated some truly fascinating
- 0:23sources for you. The landscape is just changing
- 0:25incredibly fast right now. Yeah, it really is.
- 0:28So here's our roadmap for today. We are tracking
- 0:31an entirely new direction for AI. First, we explore
- 0:35OpenAI's major pivot at the Cannes Festival.
- 0:38They are becoming a full -fledged ad tech powerhouse.
- 0:41Oh, yeah. Massive move. Next, we cover the absolute
- 0:44death of prompting. We look at the rise of loop
- 0:47engineering instead. Right. And the ecosystem
- 0:50updates supporting those autonomous agents. Exactly.
- 0:53And finally, we unpack a breakthrough in financial
- 0:55data. It completely changes how AI reads complex
- 0:58SEC filings. But I have a vulnerable admission
- 1:02before we begin. Okay, let's hear it. I still
- 1:04wrestle with prompt drift myself. You know, you
- 1:07write a prompt, but you just lose the underlying
- 1:09logic. Well, we all do. I mean, prompting is
- 1:12inherently messy. It's unpredictable. It's exactly
- 1:15why the industry is moving past it. Yeah. So
- 1:17let's start right at the commercial peak today.
- 1:20To understand where AI is actually going next,
- 1:22we have to look at how it makes money. Always
- 1:25follow the capital. It reveals the true infrastructure.
- 1:27Right. So OpenAI went to the Cannes Lions Festival
- 1:31recently, and the narrative around them has completely
- 1:34shifted this year. They aren't just a quirky
- 1:36tech lab anymore. They are arriving as a media
- 1:40and advertising company. Which is huge. Credio
- 1:42actually invited them to the Cannes Festival.
- 1:44That is a major signal to traditional advertising
- 1:47executives. Just days before Cannes, they made
- 1:50a huge move. They expanded their self -serve
- 1:52ads manager in the UK. Yeah. They officially
- 1:54rolled out cost - per click bidding. Cost per
- 1:57click means advertisers pay only when users actually
- 2:00click. It is the exact model that built modern
- 2:03search. And now it is being applied to conversational
- 2:07AI interfaces. Exactly. And OpenAI's chief revenue
- 2:11officer headlined some core sessions there. The
- 2:14main topic was advertising in the age of AI.
- 2:17Their core strategy here is becoming incredibly
- 2:20obvious. I mean, they are building infrastructure
- 2:22to monetize their massive traffic. ChatGPT generates
- 2:27billions of interactions every single month.
- 2:30Yeah, the scale is wild. And agencies are demanding
- 2:33something entirely different these days. They
- 2:35aren't just buying standalone AI models anymore.
- 2:38They want end -to -end marketing operating systems.
- 2:42Right. They want systems that run, orchestrate,
- 2:44and optimize operations. They need an engine
- 2:46that handles... So let's unpack what an end -to
- 2:50-end system actually means. It doesn't just write
- 2:52the ad copy for you. No, not at all. It targets
- 2:54the user, bids for the placement, and deploys
- 2:57it. It monitors the analytics and adjusts the
- 2:59spend automatically. It essentially replaces
- 3:01an entire floor of media buyers. Yeah. OpenAI
- 3:04is using Khan to prove a major point. They want
- 3:07to be the commercial backbone for major brands.
- 3:09It's like they stopped selling the engine entirely.
- 3:12They started selling the whole highway system
- 3:14instead. That's a perfect structural analogy.
- 3:16They were providing the roads, the signs, and
- 3:19the destination. There was also some initial
- 3:22buzz about a film project. An AI -assisted animated
- 3:25film called Critters was supposed to debut. Right,
- 3:28but that project reportedly stalled out recently.
- 3:31It was tied to the shutdown of Soar's consumer
- 3:33tool. It really shows where their true priorities
- 3:36are right now. Consumer video tools are flashy
- 3:39but highly expensive. The real story to watch
- 3:42is their ad tech infrastructure. But let me ask
- 3:44you this. Is this just about monetizing chat
- 3:48GPT traffic or s***? something bigger oh it is
- 3:51definitely something much bigger than just chat
- 3:53traffic they're building the fundamental commercial
- 3:55backbone for brands. They want to intermediate
- 3:58every single transaction on the Internet. So
- 4:01they're becoming the operating system, not just
- 4:03the tool. Exactly. And that leads us directly
- 4:05to how that system runs. Right. If OpenAI is
- 4:08building this comprehensive marketing operating
- 4:10system, how does that system actually function
- 4:13in the real world? It requires AI that doesn't
- 4:15wait for human commands. It requires agents that
- 4:18can think and act continuously. And that brings
- 4:20us to the agentic shift. We are moving away from
- 4:23traditional prompt engineering completely. Thank
- 4:26goodness. Yeah. Now, loop engineering is the
- 4:28new standard in the industry. Loop engineering
- 4:31means AI agents that keep working autonomously
- 4:34until a goal is met. Forrest Cherney from Claude
- 4:37Code recently highlighted this shift. The team
- 4:40over at OpenClaw is seeing it too. Let's clearly
- 4:42define how this actually works. You give them
- 4:44a goal, and they just loop until finished. Yeah,
- 4:47they evaluate the problem, take an action, and
- 4:49check their work. Then they loop again until
- 4:52the objective is fully met. Like Moonshot AI,
- 4:55they just added goal mode to Kimi work. It is
- 4:58a desktop agent that tracks progress continuously.
- 5:01Right. You set one objective and the agent just
- 5:04keeps grinding. You can check its progress anytime
- 5:07you want. It is a fundamental shift in how we
- 5:09interact with software. You aren't typing back
- 5:12and forth anymore. You are managing a persistent
- 5:14digital worker. But running these persistent
- 5:16agents requires serious computing power. You
- 5:20can't just leave a laptop running constantly.
- 5:22No, you really can't. That is where new tools
- 5:24like Agent 37 come in. Agent 37 offers managed
- 5:28hosting for persistent AI agents. It supports
- 5:31popular agents like Hermes, OpenClaw, and CloudCode.
- 5:34So you don't need to run them on a Mac mini.
- 5:36They live in the cloud and work around the clock.
- 5:39We are also seeing tools like Atomic Male Agentic
- 5:42emerge. It gives autonomous AI agents their own
- 5:46real inbox. Right. They manage it fully without
- 5:49human setup or ongoing intervention. They literally
- 5:52read the incoming emails, graph responses, and
- 5:54send them off. I have to push back on that specific
- 5:57idea, though. Handing over autonomous control
- 5:59of a real business inbox. Yeah. Letting an AI
- 6:02reply to sensitive emails without supervision.
- 6:05Right. Beat. That feels like a corporate disaster
- 6:07waiting to happen. It is a very valid fear to
- 6:10have right now. But the technology is adapting
- 6:12quickly to address that exact risk. That is exactly
- 6:15why tools like Backgrind are gaining traction.
- 6:18Right. Backgrind runs your AI agents over any
- 6:20app or game. But it only pings you when it actually
- 6:23needs a decision. Right. And that is the crucial
- 6:26safety mechanism we need. It handles the mundane
- 6:29routing and basic formatting tasks. But it escalates
- 6:33the high stakes choices directly to a human.
- 6:36We're also seeing this autonomy inside our daily
- 6:38documents. Grok for Word now generates outlines
- 6:41right inside the panel. Oh, wow. It can restructure
- 6:44wording and create complex tables autonomously.
- 6:47It doesn't wait for you to highlight and click
- 6:49format. And Cloudback MCP connects to clients
- 6:52like Cursor or VS Code. It manages 300 backup
- 6:55definitions through simple chat commands. So
- 6:58what happens when an agent gets stuck in a loop
- 7:00trying to solve a problem? Well, new tools are
- 7:03being built specifically to handle agent friction.
- 7:05When they get suck, they trigger a decision ping
- 7:08to you. They outline the specific roadblock and
- 7:11ask for your human intuition. Got it. Human oversight
- 7:13shifts from micromanaging to just approving decisions.
- 7:17Exactly. You become the senior editor, not the
- 7:19junior writer. Right. So to make these continuous
- 7:21loops run smoothly, things must change. The underlying
- 7:25foundation models need a major structural upgrade.
- 7:28They definitely do. They need more speed, larger
- 7:30memory, and better debugging tools. You cannot
- 7:33run persistent loops on slow, forgetful foundation
- 7:37models. They will just lose track of the original
- 7:39goal completely. Well, rumors about Cloud Sound
- 7:42at 5 are really heating up now. It is currently
- 7:44codenamed Fennec inside the development circles.
- 7:47Word is it could launch as early as next week.
- 7:50The hardware specs leaking out are highly impressive.
- 7:54Yeah. It reportedly features a 1 million token
- 7:56context window. Tokens are tiny chunks of words
- 8:00AI uses to read and write. Right. And a context
- 8:03window is the AI's active short -term memory
- 8:06limit. It is supposed to have vastly stronger
- 8:08coding capabilities, too. A million tokens means
- 8:11it remembers massive amounts of data. It can
- 8:14hold hundreds of textbooks in its active memory.
- 8:16We are also seeing Grok build remote showing
- 8:18up online. It looks unfinished and was likely
- 8:20leaked by accident. Yeah, it is unclear if it
- 8:23lives on. folds into cursor cursor is a wildly
- 8:26popular ai code editor right now a grok integration
- 8:29would disrupt that market significantly definitely
- 8:32there is also big money moving in the debugging
- 8:35space elastic is buying an ai debugging startup
- 8:38called deductive ai the deal is reportedly worth
- 8:42up to 85 million dollars deductive ai is a startup
- 8:46from 2023 with solid revenue yeah they've reached
- 8:49about 1 million in annual recurring revenue already
- 8:52This will massively boost Elastic's AI observability
- 8:56tools. When an agent hallucinates, you need to
- 8:58know exactly why. You have to trace the error
- 9:00back to its source. It's like buying an MRI machine
- 9:03specifically to diagnose where the AI went wrong.
- 9:06You need to see the exact layer inside the black
- 9:09box. That is exactly what observability tools
- 9:11provide for developers today. They map out the
- 9:14AI's internal logic pathways clearly. We are
- 9:17also seeing fascinating Y Combinator -backed
- 9:19startups emerge. Palmier just unveiled an entirely
- 9:22new video editor tool. Oh, yeah. Cloud or Codex
- 9:25can directly edit videos inside the software.
- 9:27It works seamlessly with Sedans 2 .0 and Cling
- 9:29D3. And it is completely free to download right
- 9:32now. That strategy really drives rapid user adoption
- 9:35in competitive markets. There are some quick
- 9:37hits for your toolbox today, too. Microsoft opened
- 9:41a free 12 -week AI course online. It has 24 lessons
- 9:46covering classic AI concepts thoroughly. It dives
- 9:50into deep learning, neural networks, and modern
- 9:53architectures. It is an excellent resource for
- 9:55anyone trying to catch up. Stanford also released
- 9:57a brilliant four part prompting technique. If
- 10:00your chat bot gives shallow answers, try this
- 10:03specific framework. Yeah, it pushes any model
- 10:06to give deeper Ph .D. level analysis. It is perfect
- 10:09for complex reports, interviews or big structural
- 10:12decisions. Let me ask you, why do we need a one
- 10:15million token window just for coding? Massive
- 10:18context allows the AI to ingest everything at
- 10:21once. It can literally read an entire company's
- 10:23code base simultaneously. Wow. Yeah. So it understands
- 10:26how a change here breaks a feature over there.
- 10:28You need the whole map to navigate, not just
- 10:30a single street. Precisely. And navigating that
- 10:32map requires flawlessly clean data. That brings
- 10:35us to our final highly important topic today.
- 10:37Even with huge context windows and these persistent
- 10:40working loops, the AI is completely useless if
- 10:44the data is garbled. Garbage. Always guarantees
- 10:47garbage out. Especially in highly complex systems.
- 10:51Nowhere is this more critical than in high -stakes
- 10:53global finance. A joint research team just dropped
- 10:57a monumental new fix. Yeah, they did. Teams from
- 10:59Stanford, the University of California, and Nanjing
- 11:02University collaborated. They tackled one of
- 11:04the most frustrating problems in modern AI. Feeding
- 11:07raw SEC filings into LLMs is notoriously difficult.
- 11:11LLMs are large AI systems trained on massive
- 11:15text data sets. Right. If you've tried it, you
- 11:17already know the painful truth. Financial tables
- 11:20usually get completely flattened by the AI models.
- 11:23Rows and columns just collapse into a useless
- 11:25text paragraph. Exactly. And that destroys the
- 11:27actual underlying accounting logic entirely.
- 11:30A misplaced decimal or a flattened row changes
- 11:33millions of dollars. The research team developed
- 11:36a new data set and methodology entirely. It is
- 11:38designed specifically to make SEC filings readable
- 11:41for machines. It allows the AI to parse the data
- 11:44without losing structural meaning. The researchers
- 11:46propose something they officially call SCFD.
- 11:49It is a reconstructed version of the SEC's edGR
- 11:53document filings. Right. It translates the raw
- 11:56data into layout -faithful multi -markdown. Multi
- 11:59-markdown is a text format preserving complex
- 12:02structural formatting clearly. It acts as a perfect
- 12:05bridge between human documents and machine code.
- 12:07They successfully retain the merged headers and
- 12:10the critical indentation. They keep the financial
- 12:13signs, spans, and complex table hierarchies.
- 12:16And incredibly, it uses far fewer tokens. The
- 12:24scale of this data set is just absolutely staggering.
- 12:26The team dropped 152 billion token public snapshot
- 12:31recently. Processing the full archive yields
- 12:34around 550 billion tokens total. That is 550
- 12:37billion tokens of structurally perfect financial
- 12:40documents. It is an unprecedented amount of high
- 12:43quality training material. This data set has
- 12:45less than a 0 .1 % overlap with Common Crawl.
- 12:48Common Crawl is the standard web scrape data
- 12:50most models use. Right, which means it offers
- 12:53highly unique, specialized training material.
- 12:55Whoa. Imagine scaling to a billion queries on
- 12:59flawless financial data. Two sec silence. It
- 13:02completely revolutionizes algorithmic trading
- 13:05and financial risk analysis forever. This represents
- 13:07a massive infrastructural upgrade for the financial
- 13:10AI space. It finally provides a clean, standardized
- 13:13pipeline for the industry. Millions of complex,
- 13:17unstructured financial documents become high
- 13:19-quality training data. But how does retaining
- 13:21a simple text indentation actually change the
- 13:25AI's financial logic? Well, in accounting, indentation
- 13:28signifies vital parent -child relationships and
- 13:30revenue streams. If operating income isn't indented
- 13:33under gross profit, the underlying math breaks.
- 13:36Oh, I see. The AI needs that exact visual hierarchy
- 13:39to understand the financial reality. Clean data
- 13:42finally replaces messy guesswork in financial
- 13:44AI models. It truly is the foundation for the
- 13:47next generation of financial intelligence. Mid
- 13:49-roll sponsor, BreakMarker. We have covered a
- 13:52tremendous amount of ground today. Let's pull
- 13:54this deep dive together and look at the... big
- 13:56picture. We started with OpenAI becoming a comprehensive
- 13:59ad tech operating system. They are building the
- 14:03commercial infrastructure for the modern web.
- 14:06They're moving from a simple tool to the foundational
- 14:08layer. We saw how loop engineering is fundamentally
- 14:12changing our daily workflows. It is turning AI
- 14:15from a passive assistant into a persistent worker.
- 14:18Right. We explored the rumors of robust new foundation
- 14:21models arriving soon. Models with active memory
- 14:24large enough to understand entire corporate ecosystems.
- 14:27And finally, we looked at how researchers are
- 14:30fixing raw financial data. They are repairing
- 14:32the very foundation of financial training materials
- 14:35entirely. All these pieces are snapping together
- 14:38like Lego. blocks right now. We are building
- 14:40a fully autonomous digital economy from the ground
- 14:42up. Yeah. The persistent agents, the clean data,
- 14:45the monetization framework, it is all finally
- 14:48connecting. It is a genuinely fascinating time
- 14:50to be watching this technology evolve. I want
- 14:53to leave you with one final thought today. If
- 14:55AI can run your marketing campaigns at con automatically,
- 14:58if it can manage your inbox without ever asking
- 15:01permission, and if it flawlessly parses 550 billion
- 15:05tokens of complex SEC beta, what is the deeply
- 15:08human skill you need to be cultivating tomorrow?
- 15:10Beat. Thank you for taking this deep dive with
- 15:12us today. Out to your music.