Latest / AI Fire Daily / 🎙️ EP 304: Anthropic Custom Chips with Samsung & China Hits Breakthrough with LongCat-2.0
Transcript
- 0:00We are looking at a really fractured ecosystem
- 0:02today. Whoa. I mean, imagine scaling a training
- 0:06run across 50 ,000 unproven domestic chips just
- 0:11to build one massive AI brain. Beep! That scale
- 0:15is just wild. Welcome to our deep dive. Today
- 0:19we are unpacking the global AI control race.
- 0:22There is just so much shifting right now. Yeah,
- 0:24let me lay out the roadmap for you. We are exploring
- 0:27three distinct layers of this race. First, major
- 0:30US software giants are pivoting hard. They're
- 0:33scrambling to build their own physical hardware.
- 0:35Second, we will examine the messy reality of
- 0:38software. We are looking at complex enterprise
- 0:40AI deployments. Third, we have a massive breakthrough
- 0:44from China. It proves total hardware independence
- 0:46is now possible. It is a totally different landscape.
- 0:49Exactly. Yes. And this matters directly to you.
- 0:52The people who control the physical chips dictate
- 0:54everything. They determine the capabilities of
- 0:57the AI tools you use. Right. They control the
- 0:58cost and availability, too. So let us begin at
- 1:01the foundational hardware layer. Everyone relies
- 1:03heavily on NVIDIA right now. Yeah. NVIDIA is
- 1:06definitely the undisputed king. But the software
- 1:08giants are actually rebelling. They want to control
- 1:10their own economics. Anthropic is a perfect example
- 1:13of this shift. They are in active talks with
- 1:16Samsung right now. Right, to manufacture their
- 1:18own custom AI chips. Exactly. They officially
- 1:21want a diversified hardware stack. We do not
- 1:23know the exact workloads yet. But Samsung is
- 1:26a highly logical manufacturing partner. They
- 1:28already build chips for NVIDIA, you know. And
- 1:31they were linked to Google's custom chips, too.
- 1:33So there's a total sense. Honestly, I still wrestle
- 1:36with prompt drift myself. Yeah, it happens to
- 1:38all of us. So I can only imagine the headache
- 1:41of designing custom silicon. It is just mind
- 1:44-blowing. But OpenAI just revealed a custom processor
- 1:47too. Right. They are calling it Jalapeno. They
- 1:50developed this new processor with Broadcom. And
- 1:52Broadcom is huge in networking silicon. Exactly.
- 1:55Jalapeno targets higher efficiency and performance
- 1:58per watt. It is essentially built as a custom
- 2:00inference processor. Let's define that real quick
- 2:03for everyone. That is a chip built just for running
- 2:05AI models, not training them. Yeah, spot on.
- 2:08It just drastically lowers the cost of running
- 2:10models. Amazon and Google already offer custom
- 2:14TPUs in their clouds. Right. It really reminds
- 2:17me of the restaurant business. It is like restaurants
- 2:19deciding they must farm their own ingredients.
- 2:22They just need to keep menu prices down. That
- 2:24is a really great way to put it. Two -sec silence.
- 2:27But let me push back on this a bit. Doesn't building
- 2:30custom silicon risk locking these companies into
- 2:33obsolete architectures? Well... It is a huge
- 2:36technical gamble for sure. The algorithms change
- 2:39so incredibly fast, but relying solely on NVIDIA
- 2:43is a much bigger threat. The efficiency gains
- 2:45usually outweigh the architectural risks. So
- 2:48custom chips bring efficiency but demand massive
- 2:51Risky upfront bets on architecture. Exactly.
- 2:54It is a massive financial balancing act. Let
- 2:56us pivot from theoretical hardware perfection
- 2:58now. We need to look at the software reality
- 3:00check. The reality of deploying these models
- 3:03is incredibly messy. Oh, it is way harder than
- 3:06people admit. Mark Zuckerberg actually told Metastaff
- 3:09this recently. He said AI agents are progressing
- 3:11much slower than expected. The payoff has not
- 3:14fully arrived yet. And that is despite major
- 3:16layoffs and internal reassignments. Yeah, the
- 3:19underlying models are smart. Obviously. But the
- 3:21agents themselves still struggle in the real
- 3:24world. Microsoft is realizing they have to fix
- 3:27this manually. Right. They just launched their
- 3:29new frontier company. Exactly. They put $2 .5
- 3:32billion behind it. They have 6 ,000 experts just
- 3:36to help enterprises install AI properly. Which
- 3:40brings up a big question. What is that? Well,
- 3:43if these models are so intelligent, why does
- 3:45Microsoft need an army of 6 ,000 human experts
- 3:47just to install them in businesses? Because enterprise
- 3:50data is a complete nightmare. You have decades
- 3:52of unstructured, heavily siloed proprietary data.
- 3:56Right, like old HR files or locked PDFs. Exactly.
- 3:59It is like stacking Lego blocks of data, but
- 4:01none of the pieces fit together. An AI model
- 4:04needs perfectly pristine data pipelines. Those
- 4:076 ,000 experts are basically highly paid data
- 4:10plumbers. Raw intelligence is useless without
- 4:13deep integration into messy corporate data systems.
- 4:15Exactly right. And we are seeing geopolitical
- 4:17moves here, too. OpenAI just proposed a very
- 4:19wild financial arrangement. Oh, the sovereign
- 4:22wealth fund thing. Yeah. They want to give 5
- 4:25% of their equity to a U .S. sovereign wealth
- 4:27fund. It lets the public share the upside. And
- 4:30obviously, it eases a lot of political pressure.
- 4:32It needs Congress approval. But it is a huge
- 4:35move. Meanwhile, regular developers are getting
- 4:37incredible new tools. Cloud Fable 5 came back
- 4:40yesterday morning. People are building insane
- 4:42things with it. Cloud Code also supports artifacts
- 4:45now, right? Yes. Interactive pages you can just
- 4:47share via a private link. Google also shipped
- 4:50a new AI agent stack in VS Code. That includes
- 4:53GenKit and then ADK 2 .0, right? Yep. Plus, GitHub
- 4:57co -pilot CLI now has an auto mode. It just routes
- 5:00your tasks to the best real -time model. And
- 5:03Meta released that limited app called Pocket.
- 5:05You just use a prompt to generate playable minigames.
- 5:08The consumer wins are just stacking up so fast.
- 5:10So connecting open AI's political moves to the
- 5:14global stage, we are seeing a massive shift in
- 5:17AI control. China just hit a truly massive breakthrough.
- 5:21Yeah, this is a huge turning point. Meituan,
- 5:24the Chinese food delivery giant, released Longcat
- 5:272 .0. It is a new open source coding model, and
- 5:31it is ranking in the top three globally. by call
- 5:34volume. It is available right on their site and
- 5:37open router. The specs on this thing are pretty
- 5:39wild. It is a 1 .6 trillion parameter MoE model.
- 5:44Let's clarify that term quickly. Moe just means
- 5:47a model that only uses parts of its brain for
- 5:49each task. Right. It is highly efficient. It
- 5:52has 33 billion to 56 billion active parameters,
- 5:55and it boasts a massive 1 million token context
- 5:58window. Which is the amount of text an AI can
- 6:00remember in one single conversation. Exactly.
- 6:03But the crazy part is how they built it. Right.
- 6:05It was trained entirely from scratch on domestic
- 6:07chips. They used a cluster of 50 ,000 Chinese
- 6:10chips. And they used a domestic alternative to
- 6:12NVIDIA's NCCLC. software. That software is critical
- 6:15to actually stabilize the training run. We should
- 6:18frame the context here purely on the facts. In
- 6:202022, Washington implemented export controls
- 6:23on advanced semiconductors. Since then, Chinese
- 6:25tech giants have actively sought self -reliance.
- 6:28DeepSeek and Alibaba used domestic chips for
- 6:30inference earlier. But Meijuan just proved domestic
- 6:33superpods can handle pre -training. And pre -training
- 6:36is the incredibly demanding phase, right? Yes,
- 6:39it requires perfectly synchronized data across
- 6:41tens of thousands of chips. It is incredibly
- 6:44difficult to stabilize. So looking at this timeline,
- 6:47does this mean that 2022 U .S. export controls
- 6:50essentially forced China to accelerate the maturity
- 6:54of their own hardware ecosystem? Structurally
- 6:57speaking, yes. The constraints cut off external
- 6:59supply chains. So funding and developing domestic
- 7:03innovation became an absolute necessity for survival.
- 7:06So the restrictions essentially acted as a catalyst
- 7:08for their own domestic compute evolution. Precisely.
- 7:11They had no choice but to build it themselves.
- 7:13Two sec silence. It is time to step back and
- 7:16look at the big idea here. We are witnessing
- 7:18a great decoupling at literally every level.
- 7:21We really are. The software companies are decoupling
- 7:23from NVIDIA right now. You see it with Jalapeno
- 7:25and Anthropic working with Samsung. And entire
- 7:28nations are decoupling from foreign hardware
- 7:31supply chains. Longcat 2 .0 proves that complete
- 7:35hardware independence is totally possible. Everyone
- 7:38is racing to secure their own physical foundations.
- 7:41Which leaves you with a really provocative thought
- 7:42to mull over. If every major tech giant builds
- 7:45custom chips and every global superpower builds
- 7:48totally divergent local hardware, will the global
- 7:51Internet fragment into completely separate, incompatible
- 7:54AI ecosystems? It is a very real possibility
- 7:56now. It changes how we think about the future
- 7:58of tech. Thank you for diving deep with us today.
- 8:01Keep questioning the systems you use every day.
- 8:03Stay curious, everyone. We will see you next
- 8:05time. Out to your own music.