Latest / AI Fire Daily / 🎙️ EP 315: Researcher Challenges "New Jobs" AI Narrative & NVIDIA Drops Open RAG Models
Transcript
- 0:00For centuries, human adaptability has been our
- 0:02ultimate safety net. Beat. Technology destroys
- 0:06old jobs, sure, but it always creates new ones.
- 0:09We have historically relied on that comforting
- 0:11economic rule, but, well, that safety net might
- 0:15have just evaporated completely. Two -sec silence.
- 0:19Yeah, it is a profound and honestly pretty unsettling
- 0:22technological shift. Welcome to our deep dive
- 0:24for today. We're unpacking a massive paradigm
- 0:27shift in human work. We will also analyze the
- 0:30staggering pace of recent AI news. Right. The
- 0:34news cycle is just moving at light speed right
- 0:36now. And we will explore NVIDIA's latest breakthroughs
- 0:38in agent memory. We really want to guide you
- 0:41through this complex, shifting landscape. We
- 0:43definitely have a lot of ground to cover today.
- 0:45I mean, the daily advancements are starting to
- 0:46blur together. Let's start with the biggest conceptual
- 0:49hurdle we face. Daniel Kokotajlo is a former
- 0:53researcher at OpenAI. He recently shared some
- 0:55fascinating insights on a popular podcast. His
- 0:58perspective is incredibly sobering for anyone
- 1:00paying attention. He fundamentally challenges
- 1:02our basic assumptions about the future. The historical
- 1:05rule always felt so reliable for us. Technology
- 1:08automated narrow tasks, and humans simply shifted
- 1:12their focus. Exactly. Automation was always incredibly
- 1:15narrow in its actual scope. The mechanical loom
- 1:18automated weaving, but not complex problem solving.
- 1:22The factory assembly line automated purely physical
- 1:25repetitive movements. Humans just moved into
- 1:28brand new, more complex roles. We were always
- 1:31the ultimate reliable fallback plane. Right.
- 1:33And we possessed highly generalized learning
- 1:35capabilities. The industrial machines did not
- 1:38possess that kind of adaptability. But Kokatajl
- 1:40points out a massive, insurmountable brick wall.
- 1:44Major AI labs have one single stated goal right
- 1:48now. They aim to build systems capable of all
- 1:51cognitive labor. That completely changes the
- 1:53entire economic equation for us. Yeah, it does.
- 1:56A superintelligent system is not just another
- 1:58narrow tool. It is not just a better, faster
- 2:01mechanical loom. It is being designed to be better
- 2:03than us globally. It is designed to be infinitely
- 2:06cheaper at absolutely everything. This is where
- 2:08the concept gets really daunting. Imagine a brand
- 2:11new job category is created tomorrow. A job created
- 2:14directly by the AI revolution itself. The AI
- 2:19is already perfectly equipped to do it. It creates
- 2:21a perfectly closed, self -sustaining economic
- 2:24loop. The AI doesn't just replace the old legacy
- 2:27jobs. It preemptively absorbs the newly created
- 2:30jobs as well. It's like playing a frantic game
- 2:32of musical chairs. But now the chairs are learning
- 2:36how to play. And they're playing the game better
- 2:38than we can. That is a brilliant and terrifyingly
- 2:41accurate analogy. Beat. We always assumed our
- 2:44complex brains were the safe harbor. We thought
- 2:46cognitive flexibility was our permanent undeniable
- 2:49advantage. I mean, I still wrestle with prompt
- 2:51drift myself, honestly. The idea of a machine
- 2:54constantly out adapting us is daunting. It forces
- 2:57a total uncomfortable reevaluation of human value.
- 3:00If we cannot outthink the machine, what are we
- 3:02doing? Which brings up a really crucial philosophical
- 3:05question for you. How does this shift our fundamental
- 3:07drive to learn? It basically strips away the
- 3:10economic incentive for cognitive mastery. If
- 3:13learning no longer guarantees financial security,
- 3:16learning becomes intrinsic. Right. We will learn
- 3:18for personal fulfillment, not just market survival.
- 3:21It forces us to define human value completely
- 3:23differently. So we adapt by finding purpose outside
- 3:27of purely cognitive labor. Precisely. Our future
- 3:30purpose must transcend mere economic productivity.
- 3:34Let's connect this theoretical loss to our current
- 3:36reality. We are seeing unstoppable momentum this
- 3:40very week. Beat. The culture and the industry
- 3:43are moving incredibly fast. It is genuinely hard
- 3:46to keep up with the sheer volume. Take the viral
- 3:48Erling Haaland video from a few days ago. The
- 3:51soccer star gets scared by his own mere reflection.
- 3:54Yeah, it looked incredibly real to almost everyone
- 3:57who watched it. But it was likely a highly sophisticated
- 3:59AI -generated deepfake. You really have to watch
- 4:02both versions to decide yourself. This touches
- 4:05on a massive growing issue of public trust. Christopher
- 4:08Nolan recently weighed in on this exact cultural
- 4:10problem. He called AI an obvious, dangerous Trojan
- 4:14horse. That is a very strong phrase from a major
- 4:16director. He essentially said, everyone knows
- 4:19what is hiding inside it. He noted that young
- 4:22people are actively calling it AI slop. And he
- 4:25thinks those younger audiences might actually
- 4:26be right. We are definitely seeing that slop
- 4:29flood into app stores. Apple says nudify apps
- 4:32are strictly banned on their platform. Yet security
- 4:35officials still found eight apps by passing the
- 4:38filters. Yeah. And these specific apps were directly
- 4:40linked to harmful deepfake generation. Apple
- 4:43removed three and developers could lose their
- 4:45accounts entirely. It clearly shows how difficult
- 4:48platform enforcement has truly become. The generative
- 4:51volume is simply overwhelming the manual review
- 4:54processes. And this highlights a really important
- 4:56technical transition here. Generating infinite
- 4:59garbage content requires massive amounts of computing
- 5:01power. Right. And that slop is actively eating
- 5:04up valuable server capacity. Despite the slop,
- 5:07the global consumer demand is just insane. Right.
- 5:10Kini K3 just had to halt all their new subscriptions.
- 5:13The server demand was simply too high for their
- 5:15current infrastructure. Newcomers are now permanently
- 5:18stuck on a growing waitlist. Existing users are
- 5:21safe, but it shows the massive strain. Fable
- 5:245 is doing something very similar this week.
- 5:26Starting today, their platform is strictly limited
- 5:29to premium tiers. Pro and team standard users
- 5:33are getting heavily restricted now. They will
- 5:35receive $100 credits for paper use access. It
- 5:39is a desperate move to manage overwhelming server
- 5:41traffic. Then you look over at the massive enterprise
- 5:43side. Fireworks just raised $1 .5 billion in
- 5:48funding. They raised that at a $17 .5 billion
- 5:51valuation. That is a staggering amount of capital
- 5:54for infrastructure. They already passed $1 billion
- 5:56in annualized revenue. Their platform currently
- 5:59processes over 40 trillion tokens. every single
- 6:02day most of that data flows through highly customized
- 6:05enterprise ai models 40 trillion tokens daily
- 6:08is genuinely hard to even comprehend this massive
- 6:11boundless digital demand is creating a brutal
- 6:14squeeze we are seeing a severe bottleneck in
- 6:17physical hardware globally the current asml situation
- 6:20perfectly illustrates this physical bottleneck
- 6:22beat they are facing unprecedented pressure to
- 6:25deliver manufacturing equipment They are a massive,
- 6:28critical chip hardware giant in Europe, and they
- 6:32are offering 20 ,000 euros in shares to employees.
- 6:35They are doing this just to convince them to
- 6:37stay until 2030. Around 44 ,500 employees qualify
- 6:42for this retention bonus. They are desperately
- 6:44preparing for even stronger AI hardware demand.
- 6:48Building these EUV lithography machines is incredibly
- 6:51complex manual labor. They absolutely need elite...
- 6:55human talent to build the physical infrastructure.
- 6:57Without those specialized workers, the global
- 6:59chip supply chain collapses entirely. Meanwhile,
- 7:02massive new digital models just keep dropping
- 7:04continuously. Quinn 3 .8 Max Preview is now officially
- 7:08live for developers. You can test its capabilities
- 7:10on Quinn Chat and Alibaba Cloud. We're still
- 7:13patiently waiting to see the official benchmark
- 7:15scores, but a highly anticipated open weight
- 7:18release could be coming next. I have to ask you
- 7:20about this crazy exponential curve. Is this level
- 7:23of hypergrowth actually sustainable or will we
- 7:26hit a physical wall? We are absolutely going
- 7:28to hit hard physical constraints very soon. Global
- 7:32power grids and supply chains simply cannot scale
- 7:35infinitely. The digital software side is advancing
- 7:38much faster than physical hardware. Right. Boundless
- 7:40digital demand is bottlenecked by physical chips
- 7:43and human talent. Mid -roll sponsor break. Welcome
- 7:47back to our ongoing deep dive. Before the break,
- 7:50we discussed the intense global hardware squeeze.
- 7:53Now let's look at how software developers are
- 7:55actively adapting. They essentially have to make
- 7:58the software smarter and vastly cheaper. desperately
- 8:01need to cope with that massive, overwhelming
- 8:02server demand. NVIDIA just dropped a truly massive
- 8:06new software release. They released advanced
- 8:08open models for RAG and agent memory. These are
- 8:11highly optimized, commercially available AI embedding
- 8:14models. They're built specifically to handle
- 8:16complex agentic data retrieval. Let's clarify
- 8:18a highly technical concept for you first. What
- 8:21exactly is RAG? Giving an AI a searchable library
- 8:24so it doesn't hallucinate facts. That is a wonderfully
- 8:27concise way to explain the mechanism. Without
- 8:30RAG, an AI just guesses answers based on training
- 8:33data. With RAG, it actually looks up the correct
- 8:36facts first. When AI agents pull up the wrong
- 8:40information, things break. This new NVIDIA release
- 8:43is designed to fix that exact problem. It creates
- 8:46a much stronger grounding wire for enterprise
- 8:48applications. The flagship model in this release
- 8:51has a very long name. It is officially called
- 8:54the Nemotron 3 Embed 8. BBF16. It just proudly
- 8:59took first place on the RTB leaderboard. It hit
- 9:02an impressive 78 .5 % retrieval score. But let's
- 9:06explain what that leaderboard actually means
- 9:08for developers. The RTB leaderboard rigorously
- 9:10tests how accurately models find facts. It measures
- 9:13how reliably the system retrieves the correct
- 9:15context. When retrieval fails, the AI agents
- 9:18get terrible internal evidence. Right. And bad
- 9:21evidence inevitably leads to endless extra searches.
- 9:24The AI gets confused and just keeps querying
- 9:26the database repeatedly. That means significantly
- 9:28longer processing times for the end user. And
- 9:31it results in much higher token bills for the
- 9:33company. Accuracy is directly tied to the fundamental
- 9:36economics of AI. NVIDIA also dropped two smaller
- 9:401B variants alongside the flagship. These are
- 9:43specifically designed for cost and latency sensitive
- 9:46enterprise setups. The smaller 1B variant still
- 9:49scored 72 .4%. That is highly respectable for
- 9:53a much smaller, cheaper neural model. There is
- 9:56also a highly specialized Blackwell optimized
- 9:58version now available. It can easily handle up
- 10:01to twice the normal traffic volume. It does that
- 10:03while keeping over 99 % of baseline accuracy.
- 10:07All three models can process 32 ,000 tokens at
- 10:10once. We really need to unpack why that specific
- 10:13number matters. That massive context window is
- 10:16a profound technological upgrade. It means the
- 10:19AI can basically read an entire book instantly.
- 10:22It makes them perfect for reading incredibly
- 10:24long, complex documents. They can scan deep proprietary
- 10:27code bases or massive agent interaction histories.
- 10:30NVIDIA built these impressive models using some
- 10:32truly fascinating techniques. They utilized a
- 10:35method called structured pruning during the development.
- 10:38Let's explain structured pruning simply. It means
- 10:41trimming unnecessary connections so it runs much
- 10:44faster. Think of it exactly like pruning dead
- 10:46branches off a tree. The network becomes leaner,
- 10:49faster, and uses far less energy. They also use
- 10:52a technique called advanced teacher distillation.
- 10:54This is a smaller model learning directly from
- 10:57a massive one. The small AI basically absorbs
- 11:00the wisdom of the supercomputer. It allows for
- 11:03incredible efficiency without losing the core
- 11:06intelligence. They also gradually increase the
- 11:08training contexts over time. The sheer engineering
- 11:11brilliance behind this is truly remarkable to
- 11:14witness. Whoa! Imagine scaling to a billion queries
- 11:18without losing that precision. Better initial
- 11:20retrieval means your agents do much less busy
- 11:22work. You successfully feed the system accurate
- 11:25context on the very first try. Enterprise developers
- 11:28can mathematically expect far fewer redundant
- 11:30search loops. That results in noticeably lower
- 11:33token costs across the board. It completely changes
- 11:36the fundamental economics of running AI agents.
- 11:40Beat? We are also seeing other practical infrastructure
- 11:42tools emerging rapidly. They're all actively
- 11:45riding this new wave of software efficiency.
- 11:48The tool called Basert is a fantastic example
- 11:50of this trend. It efficiently runs local AI models
- 11:53directly on Apple silicon chips. And it impressively
- 11:56runs up to 6 .4 times faster. It is significantly
- 12:00faster than the standard Lama .cpp local framework.
- 12:04Just for clarity, Lama .cpp is basically code
- 12:07for running AI locally. BaseRoot makes that on
- 12:09-device AI incredibly fast and surprisingly easy.
- 12:12We're also seeing analytical tools like OpenSEO
- 12:15gaining traction. It offers highly affordable
- 12:18keyword research and deep competitor analysis.
- 12:20It seamlessly handles complex backlinks and site
- 12:22audits for smarter strategies. SpyCost is another
- 12:25really interesting consumer -focused tool we
- 12:27found. It meticulously tracks hidden product
- 12:29price histories across the internet. It actively
- 12:32reveals fake discounts so you know when deals
- 12:34are real. It protects consumers from deceptive
- 12:36algorithmic pricing on... major storefronts.
- 12:39And finally, a tool called Kobe offers privacy
- 12:42-friendly, cookie -less analytics. It accurately
- 12:45tracks website traffic and funnels, setting up
- 12:48in seconds. These tools all represent the crucial
- 12:50shift toward efficient AI application. So bringing
- 12:53this back to the core enterprise infrastructure
- 12:55debate, how does reliable memory change the commercial
- 12:59viability of AI agents? It transforms them from
- 13:02expensive experimental toys into highly scalable
- 13:05workers. Without memory, an agent furiously wastes
- 13:09money hallucinating and searching blindly. With
- 13:11reliable memory, it becomes an efficient, autonomous
- 13:14digital employee. So better memory means fewer
- 13:17search loops and much lower token bills. Exactly.
- 13:20It makes complex, autonomous workflows economically
- 13:22feasible for everyday businesses. Let's take
- 13:25a step back and look at the big picture here.
- 13:27We are clearly moving past AI as a simple, narrow
- 13:30tool. It is definitely no longer just a fun digital
- 13:32novelty. This is true even despite the influx
- 13:35of deepfake slop. NVIDIA is vastly improving
- 13:38the core foundational memory of AI. Global platforms
- 13:41are effortlessly processing trillions of tokens
- 13:43every single day. The technology is rapidly evolving
- 13:46into a true general purpose entity. Which brings
- 13:50us completely full circle to Daniel Kokotajlo's
- 13:53earlier warning. We are actively, deliberately
- 13:55building the replacement for all cognitive labor.
- 13:58We are doing it piece by piece, model by model.
- 14:01The historical safety net of human adaptability
- 14:03is truly fraying now. The digital tools we are
- 14:07building are literally designed to out -adapt
- 14:10us. It forces society to fundamentally reconsider
- 14:12the future of human purpose. We have to figure
- 14:15out what comes after purely cognitive work. Thank
- 14:19you so much for joining us on this deep dive
- 14:20today. Please subscribe and share this specific
- 14:23conversation with a curious friend, especially
- 14:25someone who is intensely curious about the rapidly
- 14:28approaching future. We always deeply appreciate
- 14:30you listening and continuously learning alongside
- 14:33us. We want to leave you with one lingering,
- 14:35provocative question today. A heavy thought to
- 14:37mull over as you navigate this changing world.
- 14:40Beat. If a superintelligence will eventually
- 14:42be perfectly equipped to do any new cognitive
- 14:44job its own revolution creates, will the only
- 14:47highly valued human jobs left in 2030 be the
- 14:50ones that require pure physical touch, empathy,
- 14:52and in -person human presence? Two sec silence.
- 14:55Outro music.