Latest / AI Fire Daily / 🎙️ EP 297: "Claude Tag" For Slack & Japan’s Fugu Model Defies Export Controls
Transcript
- 0:00Two sec silence. Think about the exact moment
- 0:03a tool stops being just a piece of software.
- 0:06Beat. It's the precise moment it officially becomes
- 0:09your co -worker. Yeah, like it stops waiting
- 0:11for you to press a button. It just, you know,
- 0:13starts anticipating your actual needs. Right.
- 0:15Welcome to the deep dive. We're pulling from
- 0:18Anthropic's beta release notes today, some leaked
- 0:20benchmark data from Sakana AI and recent SEC
- 0:25filings. Exactly. We're trying to figure out
- 0:27one central question here. What really happens
- 0:30when our software starts acting entirely on its
- 0:32own? We are exploring anthrax radical new persistent
- 0:36memory in slack. And we'll see how Japan is actually
- 0:39dodging international export controls with some
- 0:42really clever A .I. orchestration. Plus we're
- 0:45examining the massive physical ripples this tech
- 0:47is making globally. We're covering everything
- 0:49from billion dollar fundraisers to zero water
- 0:52cooling. The sheer pace of the infrastructure
- 0:54change is I mean it's simply staggering. It really
- 0:57is. I got to make a vulnerable admission right
- 0:59up front though. I still wrestle with prompt
- 1:02drift myself. Oh, absolutely. I think we all
- 1:04do. You know, you start a conversation asking
- 1:06an AI to code a website. By prompt 10, it's somehow
- 1:11giving you a recipe for pancakes. Right. It just
- 1:13completely loses the thread of the original parameters.
- 1:16Exactly. It completely loses the context. That
- 1:19makes our first topic today highly relatable
- 1:21to you. Before moving away from transient chatbot
- 1:25interactions entirely, we are finally moving
- 1:28toward an always -on teammate. Yeah, let's unpack
- 1:31this new development from Anthropic. They recently
- 1:34rolled out a major beta update. It's a brand
- 1:37new service called CloudTag. Right, and it acts
- 1:40as a proactive, deeply integrated AI teammate.
- 1:43It basically lives directly inside your company's
- 1:46existing Slack channels. Yeah, so it learns your
- 1:49unique company context just one message at a
- 1:52time. Previously, you could direct message Claude
- 1:54easily enough. You could route specific coding
- 1:56tasks through Claude code. But this new feature
- 1:59introduces something fundamentally different.
- 2:01Right, exactly. It introduces a powerful layer
- 2:04of persistent memory. And that memory fundamentally
- 2:06changes the whole workflow dynamic. Everyone
- 2:09in a specific channel shares a single identity
- 2:11now. They share one continuous, highly informed
- 2:14Claude identity. That means any team member can
- 2:18just jump into the channel. Right, and they can
- 2:20pick up the conversation exactly where someone
- 2:22else left off. And you don't even have to prompt
- 2:23it sometimes, right? Yeah, Claude Tag actually
- 2:26features a brand new ambient mode. Wait, really?
- 2:29So it's just watching the channel constantly?
- 2:31Exactly. It proactively jumps into chats without
- 2:34needing a specific prompt. It observes the flow
- 2:37of your team's conversation silently. So when
- 2:40you finally tag it with a specific job, it takes
- 2:43action. Right. It already has all the context
- 2:46to break a complex task down into stages, and
- 2:50it reports progress directly in threads. Plus,
- 2:53it uses admin siloed tools to carefully protect
- 2:56company operations. That is incredibly powerful.
- 2:59It totally changes the game. Using older AI feels
- 3:02like using a vending machine. You put in a coin
- 3:05and you get a snack. Every transaction is isolated.
- 3:07Yeah, completely separated from the last one.
- 3:09But CloudTag is like a co -worker who actually
- 3:11remembers the project from last week. That isolated
- 3:14transaction model is exactly the problem with
- 3:16chatbots. A transient model just forgets everything
- 3:19between isolated chats. A persistent model builds
- 3:22a working mental framework, though. Exactly.
- 3:24It converts your channel history into mathematical
- 3:27vectors. So it understands the unspoken rules
- 3:30of your team dynamically. Which leads me to ask,
- 3:34does persistent memory mean it stops hallucinating?
- 3:38Well, it drastically reduces the chances of making
- 3:41things up. Hallucinations often happen when a
- 3:44model lacks sufficient context. So without context,
- 3:47the system just tries to guess the answer probabilistically.
- 3:50Right. By anchoring it in shared channel memory,
- 3:52it stays thoroughly grounded. It references actual
- 3:55project history instead of pure probability.
- 3:58So memory anchors the AI. Reducing made -up facts
- 4:01over time. Exactly. It grounds the intelligence
- 4:04in shared reality. If Claude is mastering the
- 4:07micro level of team orchestration, let's zoom
- 4:10out a bit. Yeah, let's look at Japan's Sakana
- 4:12AI. They're using orchestration at the macro
- 4:14level right now. And they're actually using it
- 4:16to dodge international law. It's a wild geopolitical
- 4:19development in the AI space. Sakana AI just dropped
- 4:22a new model called Fugu. It's taking a very different
- 4:25path to frontier -level performance. Yeah, Fugu
- 4:28acts entirely as a multi -agent orchestrator.
- 4:32Let's define that jargon for the listener really
- 4:34quickly. An AI manager that delegates tasks to
- 4:37specialized helper AI programs. Perfect definition.
- 4:41The core Fugu model basically operates like a
- 4:44project manager. And Sakana is pitching this
- 4:46setup as a strategic hedge. It's a hedge against
- 4:49recent international export controls. Because
- 4:52those controls recently blocked access to major
- 4:55top -tier models, they completely pulled the
- 4:58plug on Anthropix, Mythos, and Fable in Japan.
- 5:01So Sakana built a system to route queries internally
- 5:04instead. Right. It evaluates the user prompt
- 5:07and selects smaller helper models. Then it merges
- 5:10the final answers from these various open -source
- 5:12models. And it keeps the complex multi -agent
- 5:14process completely hidden from you. incredibly
- 5:18clean API in your screen. The launch actually
- 5:20includes two distinct versions of the system.
- 5:22There's a faster, standard version called Simply
- 5:24Fugu. That was designed specifically for everyday
- 5:27coding and chat tasks, right? Exactly. Then there
- 5:30is a much heavier version called Fugu Ultra.
- 5:33That one is built for highly intensive reasoning
- 5:35tasks. And apparently both models perform extremely
- 5:38well on synthetic benchmarks. They perform near
- 5:41or above Fable 5 and Mythos. Yeah, they score
- 5:44highly across complex coding and science evaluations.
- 5:48Sakana's main selling point is pure geopolitical
- 5:51resilience. They're actively positioning Fugu
- 5:53as a highly secure alternative. It's a way to
- 5:57secure frontier level capabilities internally.
- 6:00It completely removes the risk of losing sudden
- 6:02API access. Right. You're fully protected from
- 6:05international export bans. You're no longer reliant
- 6:08on a single American tech company. I was looking
- 6:10at the early community reception, though. Users
- 6:12are reporting the real -world output is seriously
- 6:15struggling. Yeah, real -world output isn't hitting
- 6:17frontier level yet. It sounds like stacking Lego
- 6:20blocks of data. It's clever, but structurally
- 6:23wobbly. Benchmarks often fail to capture real
- 6:26-world workflow nuance. Orchestrating multiple
- 6:28weaker models introduces a heavy coordination
- 6:31tax. Context just gets lost during the handoff
- 6:34between the different agents. Exactly. The final
- 6:37output often feels somewhat disjointed and confusing.
- 6:40So, can combining weaker models really outsmart
- 6:44a massive export ban? Legally speaking, yes,
- 6:47it completely bypasses the current restrictions.
- 6:50Technically speaking, it's a much harder process
- 6:52to execute smoothly. The system struggles with
- 6:55maintaining a cohesive logical thread. It's a
- 6:57brilliant legal workaround, but it really lacks
- 7:00refinement. Clever architecture doesn't automatically
- 7:03equal frontier -level real -world intelligence
- 7:06yet. To sex silence. It's a step, but certainly
- 7:09not the destination. We're going to take a very
- 7:11quick break right here. But stick around. We've
- 7:13got a lot more to cover. Two sec silence. And
- 7:17we are back. Creating these massive orchestrators
- 7:20requires staggering amounts of capital. Oh, absolutely.
- 7:22It requires unbelievable hardware and massive
- 7:25human restructuring. We're seeing the economic
- 7:27infrastructure completely transform overnight.
- 7:29For instance, Oracle recently cut 21 ,000 internal
- 7:33jobs. Wow. That's nearly 13 % of their overall
- 7:36workforce. It's a massive, painful corporate
- 7:39restructuring. And the company stated AI deployment
- 7:42directly led to these reductions. Let's explain
- 7:45how that actually happens internally. AI automates
- 7:49the highly repetitive back office data flows.
- 7:52Right. You don't need thousands of routine administrative
- 7:54processors anymore. But we're also seeing an
- 7:57immediate legislative response to this. Yeah.
- 7:59California Representative Sam Liccardo introduced
- 8:02a major new bill. It offers up to $5 ,000 per
- 8:06impacted worker. This money flows directly through
- 8:09corporate tax credits. It's specifically for
- 8:11retraining, apprenticeships, and hiring recent
- 8:14graduates. Then you have to look at the sheer
- 8:16capital involved globally. SpaceX just launched
- 8:19a $25 billion notes offering. Wait, $25 billion
- 8:23with a B? Yes, $25 billion. They're aggressively
- 8:26using this to fund massive AI expansion. The
- 8:30deal reportedly drew nearly $85 billion in orders.
- 8:33Investors are absolutely desperate to back this
- 8:35physical infrastructure. Menlo Ventures just
- 8:37raised a new $3 billion fund. Their earlier $750
- 8:40million anthropic bet definitely paid off. That
- 8:45initial stake is now reportedly worth $14 billion.
- 8:48It proves AI still drives absolutely huge venture
- 8:51capital wins. But all this capital is paying
- 8:54for massive physical hardware. NVIDIA's zero
- 8:58-water AI factory design is getting serious attention.
- 9:01Traditional data centers evaporate millions of
- 9:04gallons of water constantly. They use it to cool
- 9:06down the massive server racks. Yeah, because
- 9:08these new chips run incredibly hot physically.
- 9:11Water absorbs heat much faster than traditional
- 9:14air cooling. But NVIDIA is using a totally closed
- 9:17-loop cooling system now. It cycles the exact
- 9:20same liquid refrigerant continuously. It could
- 9:22reduce water use to near zero globally. This
- 9:26works exceptionally well in highly favorable
- 9:28global climates. The physical scale of these
- 9:30data centers is terrifying, really. They consume
- 9:33massive amounts of raw power and resources. Whoa.
- 9:36Imagine scaling to a billion queries without
- 9:38a drop of water. It's a remarkable feat of modern
- 9:41industrial engineering. And the human talent
- 9:44required is just as intensely focused. OpenAI
- 9:48DevDay is officially taking applications right
- 9:50now. It happens in San Francisco on September
- 9:5329th. Getting a role at these core companies
- 9:55is notoriously brutal. A recent Business Insider
- 9:59piece highlighted the culture at Google DeepMind.
- 10:02Yeah, it explained the intense work ethic Vlad
- 10:04Feinberg actively used. It requires extreme dedication
- 10:07to reach that top tier. They specifically model
- 10:10hiring on his intense dedication. So with all
- 10:13this shift, it begs a question. Right. Are we
- 10:16building jobs faster than Oracle is cutting them?
- 10:18The current data suggests a very painful transition
- 10:21period. We're destroying legacy administrative
- 10:24and basic coding jobs quickly. But we're creating
- 10:27highly specialized infrastructure and complex
- 10:30orchestration roles. The overall number might
- 10:32balance out eventually over time, but the required
- 10:34skill sets are entirely different now. We are
- 10:37shifting human capital, not just eliminating
- 10:39it entirely. Exactly. It's a brutal, rapid skills
- 10:43transition. Let's look at how this trickles down
- 10:45to the individual. All these macro investments.
- 10:48hit the listener's desk eventually. Oh, they're
- 10:51hitting your everyday tools very rapidly right
- 10:53now. For example, Meta is designing its own cheaper
- 10:56AI smart glasses. They're starting at just $299.
- 11:00They use internal cameras and the Muse Spark
- 11:04AI. They can instantly identify physical objects
- 11:07around you. And they answer spoken questions
- 11:10about your immediate physical environment. It's
- 11:12basically taking the persistent memory of Claude
- 11:15Tagg. but you're strapping it directly to your
- 11:17eyeballs. Yeah. That's exactly what's happening.
- 11:20Let's run through some other rapid -fire everyday
- 11:22tools. Let's group them by the ecosystems they
- 11:25are actively building. Okay, first we have the
- 11:27entirely private AI ecosystem. Little Bird builds
- 11:30a completely private AI memory system locally.
- 11:33It silently pulls from your Slack, email, and
- 11:36daily meetings. It reads your private documents
- 11:38and your screen text. You get instant answers
- 11:41with the actual receipts attached. Yeah. And
- 11:43your private data stays entirely on your local
- 11:45system. Then there's Cotopist for the Mac operating
- 11:47system. It's a smart autocomplete tool that learns
- 11:50your habits. Right. It works seamlessly inside
- 11:52mail, Slack, and your notes. It even autocompletes
- 11:56your other complex AI prompts locally. Then we
- 11:59pivot to the massive new verification ecosystem.
- 12:02In truth, it is a wildly powerful new Chrome
- 12:07extension. This one is particularly relevant
- 12:09for the intense political sphere. It fact checks
- 12:12politicians in absolute real time. You can instantly
- 12:15verify claims during live debates or interviews.
- 12:18It actually verifies mathematical claims while
- 12:21the video is streaming. Fablewatch is another
- 12:23interesting new monitoring tool. It watches the
- 12:26Claude Fable 5 API continuously. The team automatically
- 12:29pings the server every 60 seconds, exactly. And
- 12:32the instant it answers, Fablewatch automatically
- 12:34alerts you. The visual generation side is also
- 12:37advancing incredibly fast. Yeah. Open art generates
- 12:40consistent cinematic AI videos up to five minutes
- 12:43long. It maintains consistent characters and
- 12:45scenes throughout the render. The visual style
- 12:48and voice remain completely stable automatically.
- 12:51You just type a prompt and the entire scene renders.
- 12:53With real time fact checking and instant video,
- 12:56what happens to trust? Trust is absolutely no
- 12:59longer implicit in the media. You can't trust
- 13:01a video just because you see it. You have to
- 13:03verify the cryptographic origin of the file.
- 13:05Cryptographic origins use deeply embedded mathematical
- 13:08hashes. So if one pixel changes, the entire hash
- 13:12breaks instantly. Exactly. Tools like Intruth
- 13:15will become mandatory cognitive filters. Seeing
- 13:19isn't believing anymore. Verifying the data source
- 13:22is everything. Pull all these threads together
- 13:24for a second. We're no longer just prompting
- 13:27isolated AI text boxes. We're rapidly entering
- 13:29an era of ambient persistent AI ecosystems. It
- 13:33might be an orchestrator dynamically dodging
- 13:36international export laws. It might be a slack
- 13:38bot quietly anticipating your workflow. Or it
- 13:40could be a $25 billion infrastructure push. AI
- 13:44is permanently moving from being a simple discrete
- 13:46tool. It's becoming an environment we actually
- 13:48live and work inside of. That brings us to a
- 13:51final measured question. I want you to think
- 13:53about your own role in this new ecosystem. It's
- 13:56crucial to reflect on where we fit in. If AI
- 13:58is now capable of acting as a proactive colleague,
- 14:01and if it acts as a manager of other models entirely,
- 14:05when do we stop being the primary operators?
- 14:08And when do we start becoming the ones being
- 14:11orchestrated? That is the ultimate question for
- 14:14the next decade. Look closely at your own daily
- 14:17workflows tomorrow morning. Question, which specific
- 14:19tasks are ripe for a multi -agent orchestrator?
- 14:23It's definitely going to change how you view
- 14:24your job entirely. Thank you for joining us on
- 14:26this deep dive today. We'll catch you on the
- 14:28next one.