Latest / AI Fire Daily / 🎙️ EP 278: Huawei’s Radical "Tau Scaling" Breakthrough & The Silent AI Resume Rejection
Transcript
- 0:00Imagine applying for a challenging new job today.
- 0:02You put your entire history into a resume. You
- 0:05submit it. Then you get rejected by a cold calculating
- 0:08machine. Beat. But here is the part that keeps
- 0:12me up at night. You suddenly realize that single
- 0:15algorithm just secretly blacklisted you everywhere.
- 0:18It blocked you from a hundred completely different
- 0:21companies. And no human being ever knew it happened.
- 0:24It is a profoundly chilling thought. Especially
- 0:27when you realize it is already happening. Welcome
- 0:29to today's deep dive. I am really glad you're
- 0:32joining us for this one. We have got a fascinating
- 0:34stack of research to unpack today. We really
- 0:36do. Today, we're exploring the extreme scaling
- 0:39of the new tech landscape. We will start with
- 0:42how physical hardware is being radically re -engineered.
- 0:44Then we will move into the mind -bending scale
- 0:46of modern AI software. And finally, we will look
- 0:49at the systemic dangers when these massive systems
- 0:52make decisions about human lives. I want to start
- 0:55with the physical layer. Because before AI can
- 0:58scale globally, it needs physical pathways. And
- 1:02geopolitics has forced some massive, almost desperate
- 1:05hardware innovation recently, specifically with
- 1:08Huawei. They're essentially bypassing the classic
- 1:11Moore's law. Great. For decades, Moore's Law
- 1:14was the golden rule. It basically states that
- 1:16computer processing power doubles every two years.
- 1:19It is mostly about shrinking the actual transistor
- 1:21size. But Huawei is in a very unique position.
- 1:24They are completely cut off from TSMC. TSMC being
- 1:27the world's top chip manufacturer. Exactly. Huawei
- 1:31physically lacks the advanced lithography tools
- 1:33to print those microscopic circuits. So they
- 1:35cannot just make the tiny brains of the chip
- 1:38smaller. They had to rethink the entire physical
- 1:40architecture from scratch. beat. They are introducing
- 1:44something called the tau scaling law. A lot of
- 1:46folks in the industry are actually nicknaming
- 1:48it Herr's law. And the core idea here is brilliant
- 1:51in its simplicity. If you cannot shrink the transistors,
- 1:54you shrink the time it takes for electrical signals
- 1:55to travel between them. They are focusing entirely
- 1:58on the pathways inside the machine. Let's unpack
- 2:01how they actually do that. Because they use this
- 2:03clever trick called logic folding. It is a highly
- 2:06specific circuit level design hack. Essentially,
- 2:10they fold in place related logic blocks much
- 2:14closer together. Yeah, it is a total spatial
- 2:16redesign. This approach massively reduces internal
- 2:20electrical resistance. It also drops what engineers
- 2:23call parasitic capacitance. That is basically
- 2:26unwanted energy loss between closely packed wires.
- 2:29Which is a huge issue at that microscopic scale.
- 2:31Oh, absolutely. By folding the logic, they just
- 2:33make the critical internal wires much shorter.
- 2:35I loved the analogy in our research for this.
- 2:38It's like, if we can't make the cars any smaller,
- 2:40let's just make the highway perfectly straight
- 2:41two sec silence. But I am trying to visualize
- 2:44this practically. Does shrinking the wires truly
- 2:47compensate for having physically larger transistors?
- 2:50Well, it doesn't entirely replace raw, pure density.
- 2:53You still want smaller transistors if you can
- 2:55get them. But it is an incredibly clever survival
- 2:57tactic. It focuses purely on signal speed and
- 3:00extreme efficiency. By optimizing the commute,
- 3:03you make the whole city run faster. So it's about
- 3:05signal speed, not just raw transistor size. Yeah.
- 3:08Beat. Exactly. And it is a total paradigm shift
- 3:11for them. Their goal is to reach 1 .4 nanometer
- 3:15class. density by 2031. And they claim they have
- 3:17already built 381 chips using this exact thinking.
- 3:21Being forced into a corner really created a totally
- 3:24new pathway. They're putting this in their next
- 3:26Kirin phone chip. It will be the first full test
- 3:29in the wild. I am going to be watching those
- 3:31benchmark tests very closely. It is a massive
- 3:34gamble for their engineering team. Whether this
- 3:36becomes an industry standard remains to be seen.
- 3:39But it is an amazing response to intense geopolitical
- 3:42pressure. While Huawei reorganizes those physical
- 3:45pathways, something else is happening on top
- 3:48of it. The software running on global hardware
- 3:50is scaling at a speed that is... Honestly, incomprehensible.
- 3:54The financial and usage numbers alone are just
- 3:56staggering. Take OpenRouter, for example. It
- 3:58is a major AI model hub. They just raised $113
- 4:01million. They exploded in popularity recently.
- 4:05They now serve around 8 million developers. And
- 4:07the volume of data they handle is what really
- 4:10stopped me in my tracks. Their single API processes
- 4:13an incredible 1 .5 quadrillion AI tokens annually.
- 4:17Just to clarify for everyone, a token is just
- 4:19a basic chunk of a processed word. But still.
- 4:22Whoa. Imagine scaling to 1 .5 quadrillion tokens.
- 4:26The sheer infrastructure required to route that
- 4:28traffic is insane. It is hard to even wrap your
- 4:31head around that volume. And the foundational
- 4:33models driving that traffic are evolving rapidly,
- 4:35too. Just look at Quinn. Their smartest model,
- 4:38QN 3 .7 Max, just officially landed in their
- 4:41Go plan. Yes, and its capabilities are wild.
- 4:44It effortlessly handles 2 ,000 -page contexts
- 4:47in a single prompt. More impressively, it can
- 4:49run autonomous agents for 35 hours straight.
- 4:52I still wrestle with prompt drift myself. Two
- 4:54-sec silence. You know, when the AI just forgets
- 4:56its original instructions after 10 minutes. It
- 4:58is wild to think about a system maintaining pure
- 5:00focus for 35 hours. It even plugs directly into
- 5:03the Cloud Code ecosystem now. we are also seeing
- 5:06massive leaps in abstract reasoning anthropics
- 5:09researcher sholto douglas revealed something
- 5:11huge recently their mythos model successfully
- 5:14solved the famous air doom problem number 90.
- 5:17right it achieved the exact same mathematical
- 5:19result as gbt 5 .5 but it did so with this remarkably
- 5:24cute simple proof It wasn't just brute force
- 5:27calculation. The math capabilities are reaching
- 5:30a new tier of actual reasoning. And startups
- 5:32are heavily capitalizing on this momentum. There's
- 5:35a company called Trajectory that just raised
- 5:37$15 million. It was founded by former researchers
- 5:40from OpenAI, Apple, and Google DeepMind. Their
- 5:43primary goal connects right back to that idea
- 5:45of reasoning. They want to build AI that improves
- 5:48itself from real user feedback. The distance
- 5:51between a user correcting the model and the model
- 5:53permanently improving is shrinking fast. We are
- 5:55also seeing this intelligence bleed into the
- 5:57physical world. Figure AI just signed a massive
- 6:00deployment deal. They are putting humanoid robots
- 6:02inside Catalyst Brand's logistics network. This
- 6:05brings autonomous agents directly into the physical
- 6:08supply chain. The source notes JCPenney warehouses
- 6:11might be the next target. Brooks Brothers warehouses
- 6:14are also on that potential deployment list. It
- 6:16really shows how rapidly software is automating
- 6:18the physical world. But meanwhile, on the consumer
- 6:21side, companies are locking these powerful features
- 6:24behind paywalls. Meta now has paid plans across
- 6:27their entire social ecosystem. If you use Instagram,
- 6:30Facebook or WhatsApp, they're all getting premium
- 6:33AI access. These paid tiers unlock extra creator
- 6:37tools. You get custom profiles, advanced analytics
- 6:39and deeper generative features. It is deeply
- 6:42integrating AI into daily social media habits.
- 6:45And specific utility tools are multiplying rapidly
- 6:48as well. 2 .1 is a great example of this trend.
- 6:51It is fascinating workflow. It records conversations
- 6:53directly from your personal Apple Watch. Then
- 6:56it syncs them with clogged through the MCP protocol.
- 6:58MCP is a bridge connecting AI to external tools.
- 7:01It is totally seamless data capture right from
- 7:03your wrist. Powerbase is another powerful new
- 7:05tool hitting the market. It basically builds
- 7:07AI apps using Postgres databases and RREG. RREG
- 7:11gives an AI a custom memory bank to read from.
- 7:14Exactly. Powerbase handles that memory, the workflows,
- 7:17and the automation permission. in one single
- 7:19platform. It's a very clean setup. Another integration
- 7:22tool I found fascinating is a platform called
- 7:24Zero .xyz. It gives your AI agent massive external
- 7:28access capabilities. It connects to roughly 8
- 7:31,000 different tools and APIs. Yeah, it can securely
- 7:35access services like Cloud Code and Codex. It
- 7:37links up seamlessly with Gemini. It even talks
- 7:40to CLI agents. CLI agents are text -based command
- 7:44line tools. With all this data flowing, There
- 7:46is also a strong push for user privacy. Oasis
- 7:49browser caught my eye. It is a privacy -first
- 7:52AI browser that you can train locally. You train
- 7:55it anonymously, ensuring your personal data remains
- 7:58on your machine. But amid all this breathtaking
- 7:59scale, there is a weird reality check. There
- 8:02are still glaring, almost comical flaws in these
- 8:05systems. For instance, Google's AI overview still
- 8:07cannot reliably count letters in simple words.
- 8:10There is this weird paradox I cannot get over.
- 8:13Beat. It's actually quite funny to look at the
- 8:15examples. At one point, Google's AI claimed the
- 8:19word Google has two P's. It also spelled Trump
- 8:21as T -R -P -U -M. And people just found a new
- 8:26obvious phrase that exposes AI writing instantly.
- 8:29It isn't the em dash, as many folks suspected.
- 8:32Right. And it just shatters the illusion of intelligence
- 8:34for a second. It really does. Exactly. How can
- 8:37we have models solving Erdo's math problems while
- 8:40failing to count letters in a basic word? It
- 8:43comes down to how they see the world. Complex
- 8:45reasoning patterns are vastly different from
- 8:47basic token -level spelling mechanics. They process
- 8:50abstract mathematical concepts beautifully, but
- 8:52they don't actually see individual alphabetical
- 8:54characters the way we do. Right. Massive logical
- 8:57reasoning doesn't mean it understands basic spelling.
- 8:59Beat. Precisely. It is a fundamental quirk of
- 9:02how neural networks parse information. They predict
- 9:04chunks of meaning, not individual letters. We're
- 9:07going to pause right here. We will continue our
- 9:09deep dive right after this. Sponsor. So we just
- 9:13talked about AI spelling Google with two Ps.
- 9:15The quirky unreliability of sculling mistakes
- 9:18stops being funny eventually. It gets incredibly
- 9:20serious when these shared algorithms start controlling
- 9:23human livelihoods. And that brings us to a massive,
- 9:26highly concerning Stanford research study. Letting
- 9:30an AI scan resume sounds like the ultimate corporate
- 9:32time saver, right? But this extensive new study
- 9:36showed the hidden cost of that efficiency. The
- 9:38scale of this study is huge. They reviewed 4
- 9:41million job applications in total. These applications
- 9:44spanned across 156 different employers globally.
- 9:47All of these employers were using the exact same
- 9:50AI screening tool. It is called Pometrix. The
- 9:52granular findings showed a very clear and alarming
- 9:55negative bias in the screening. Around 10 .6
- 9:58% of the job positions showed a measurable bias
- 10:00against Black applicants. They also found a 5
- 10:02.3 % bias against Asian applicants. Just the
- 10:06sheer scale of automated rejection at that level
- 10:08is deeply troubling. But the real danger, the
- 10:11thing that makes this a systemic issue, is the
- 10:13hidden cascading effect. 42 of these AI models
- 10:17were shared across different employers. Getting
- 10:19rejected by one algorithm essentially triggered
- 10:22rejections everywhere else. It affected completely
- 10:25different companies using the exact same underlying
- 10:28system. You applied to company A, got silently
- 10:31flagged, and company B rejected you too. Roughly
- 10:344 % of candidates in the study applied to 10
- 10:37different positions. They were systematically
- 10:39rejected from all of them by the machine. That
- 10:42is a much higher fail rate than human recruiters
- 10:44produce. Because human recruiters, for all their
- 10:47flaws, make independent choices. They have varying
- 10:49personal preferences and blind spots. Exactly.
- 10:52An algorithm applies the exact same rigid filter
- 10:54every single time. It creates an invisible, impenetrable
- 10:57wall for certain applicants across the entire
- 10:59job market. Now, I do want to push back on one
- 11:01thing here. We should note that this specific
- 11:03data has a time frame. The researchers rightly
- 11:06point out the data is from 2018 to 2022. Today's
- 11:09hiring wave is increasingly powered by modern
- 11:11LLMs. Are these older pieometrics findings still
- 11:15relevant today? That is a very fair pushback
- 11:17and an important contextual detail. Modern language
- 11:21models absolutely function differently than older
- 11:23predictive tools. They are much more nuanced,
- 11:25but the core structural danger hasn't changed
- 11:28at all. Because of the centralization. Yes, exactly.
- 11:31When 50 different companies rely on the exact
- 11:34same third -party AI vendor to screen resumes,
- 11:37they inherit the exact same blind spots. The
- 11:40specific biases might shift, but the cascading
- 11:43effect remains. That is why keeping a human in
- 11:45the loop is absolutely mandatory. But is the
- 11:48centralization of AI models the real enemy of
- 11:51fair hiring here? I would say yes. Identical
- 11:54shared blind spots create systemic walls for
- 11:56applicants. You cannot outsource your entire
- 11:58judgment process to a monopoly. Basically, never
- 12:01let one single AI vendor be the only gatekeeper.
- 12:04Beat. Precisely. The underlying technology evolves
- 12:07and gets smarter, but the centralization risk
- 12:09remains completely constant. Let's take a moment
- 12:11to step back and look at the whole picture. Two
- 12:14sec silence. We are looking at a complete re
- 12:17-architecture of modern technology right now.
- 12:19On the physical side, we are literally folding
- 12:22logic circuits to outsmart the limits of physics.
- 12:25And on the software side, we are processing quadrillions
- 12:27of tokens to solve incredibly complex math. We
- 12:30are sending autonomous humanoid robots into major
- 12:33retail logistics networks to move physical boxes.
- 12:36But as we hand over the keys to these shared
- 12:38algorithms, we risk building massive invisible
- 12:41walls. Right. Whether it's an AI that cannot
- 12:43spell a simple word or a resume screener locking
- 12:46people out of the economy. The overarching theme
- 12:48today is the inherent tension between breathtaking
- 12:50scale and the absolute necessity for human oversight.
- 12:54It is a delicate balancing act we are all navigating
- 12:56right now. Scale brings power, but centralization
- 12:59brings systemic vulnerability. That leaves us
- 13:02with a final thought for you to mull over. If
- 13:04AI models inherit blind spots from their centralized
- 13:07vendors. And we are moving toward a world where
- 13:09platforms like Zero .xyz let thousands of autonomous
- 13:13tools and agents talk directly to each other.
- 13:16Who is actually auditing the conversations between
- 13:19the bots when a human is no longer in the loop?
- 13:21Beat. it is the big question i strongly encourage
- 13:25you to keep questioning the automated systems
- 13:27you interact with daily thank you for joining
- 13:30us on this deep dive we will see you next time