State of AI in 2026: LLMs, Coding, Scaling Laws, China, GPUs, AGI
Guests: Nathan Lambert (Allen Institute for AI) & Sebastian Raschka (ML Researcher, Author)
Quick Take
A technical deep-dive into the current AI landscape from two respected ML researchers who actually build these systems. Unlike CEO interviews with obvious conflicts of interest, Lambert and Raschka offer practitioner perspectives on what's real vs. hype. The conversation is refreshingly honest about uncertainty — nobody's trying to sell you anything.
Key Claims Examined
🌐 "No Company Will Have Exclusive Technology"
"I don't think nowadays, in 2026, that there will be any company having access to a technology that no other company has access to. Researchers are frequently changing jobs, changing labs. They rotate. So I don't think there will be a clear winner in terms of technology access."
Our Analysis
Sebastian Raschka argues the differentiator isn't ideas — it's resources. This is a nuanced and largely accurate take:
- Talent mobility is real: Key researchers have moved between OpenAI, Anthropic, Google, and academia. Ilya Sutskever's departure from OpenAI is a prominent example. Information diffuses.
- Papers matter: Even "closed" labs publish enough research that core techniques become industry knowledge within months. Transformer architecture itself came from Google but powers everyone's models.
- The compute caveat: While ideas spread, the ability to train at scale ($100M+ runs) remains concentrated among ~10 organizations globally. Ideas are necessary but not sufficient.
- China's open-weight strategy: DeepSeek, Qwen, and others publishing detailed technical reports accelerates knowledge diffusion even further.
Verdict: Accurate — ideas diffuse, but resources concentrate
🇨🇳 Chinese Open Models as Influence Strategy
"A lot of top US tech companies won't pay for an API subscription to Chinese companies for security concerns... people at these companies see open weight models as an ability to influence and take part in a huge growing AI expenditure market in the US."
Our Analysis
Nathan Lambert frames Chinese open-weight releases as strategic, not altruistic. This interpretation holds up well:
- Security reality: Major US enterprises genuinely won't send data to Chinese API endpoints. Export controls reinforce this separation.
- Open weights bypass restrictions: If DeepSeek or Qwen is running locally in a US data center, security concerns about data exfiltration largely evaporate.
- Distribution wins: Every developer fine-tuning DeepSeek builds familiarity and ecosystem lock-in — even if no money changes hands.
- The licensing angle: Chinese models often have more permissive licenses than Meta's Llama. Less friction = more adoption.
Verdict: Plausible strategic interpretation
🤖 "Claude Opus 4.5 Hype is Almost a Meme"
"The hype over Anthropic's Claude Opus 4.5 model has been absolutely insane. I've used it and built stuff in the last few weeks, and it's almost gotten to the point where it feels like a bit of a meme in terms of the hype."
Our Analysis
This observation about X/Twitter echo chambers vs. actual usage is important:
- The X bubble: AI Twitter is dominated by developers and researchers who care intensely about coding performance. Claude Code's release aligned perfectly with this audience.
- The usage reality: ChatGPT and Gemini serve vastly more general users asking everyday questions. Their moats are distribution and brand recognition.
- Claude Code is legitimately good: Both guests acknowledge it's "way better" for coding tasks than competitors. The hype has a real foundation — it's just not representative of the broader market.
- Muscle memory matters: Sebastian's point about ChatGPT being habitual for most users mirrors how browser preferences persist despite alternatives.
Verdict: Valid critique of sample bias in AI discourse
💻 Google's TPU Advantage Over NVIDIA
"The margin on NVIDIA chips is insane and Google can develop everything from top to bottom to fit their stack and not have to pay this margin, and they've had a head start in building data centers."
Our Analysis
This claim about compute economics deserves scrutiny:
- NVIDIA margins are real: NVIDIA's data center gross margins exceed 75%. This is exceptional and gives vertically integrated competitors an opportunity.
- TPU advantage is context-dependent: Google's TPUs excel at inference and specific training workloads. For general flexibility and ecosystem compatibility, NVIDIA H100s/H200s remain the default choice for most organizations.
- The integration tax: Google's TPUs only work within Google Cloud. If you're not already in GCP, switching costs are enormous. This limits the competitive impact.
- Data center lead time: Google has been building power infrastructure and cooling systems for decades. This operational advantage is real and hard to replicate.
Verdict: Partially true — real cost advantage but limited ecosystem applicability
🔌 "OpenAI is GPU Deprived"
"OpenAI is so GPU deprived; they're at the limits of the GPUs. Whenever they make a release, they're always talking about how their GPUs are hurting... Sam Altman said, 'We're releasing this because we can use your GPUs.'"
Our Analysis
This reveals an important constraint on frontier AI development:
- The bottleneck is real: OpenAI's compute demand outpaces even their massive Microsoft Azure allocation. Inference alone for ChatGPT requires extraordinary resources.
- The open-source logic: Releasing gpt-oss means users run inference on their own hardware, reducing OpenAI's operational costs while maintaining mindshare.
- Supply chain context: NVIDIA can't manufacture chips fast enough to meet demand. Even well-funded labs face 6-12 month wait times for major GPU orders.
- The Stargate question: OpenAI's $500B Stargate announcement reflects their compute hunger. Whether it materializes is separate from whether the demand is genuine.
Verdict: Accurate — compute scarcity is a real constraint
🛠️ Tool Use as Hallucination Solution
"One of the best ways to solve hallucinations is to not try to always remember information or make things up. For math, why not use a calculator app or Python? If I ask 'Who won the soccer World Cup in 1998?' instead of just trying to memorize, it could go do a search."
Our Analysis
Sebastian's framing of tool use as an architectural solution to hallucination is insightful:
- The core insight is sound: LLMs are pattern matchers, not knowledge databases. Offloading factual lookups to reliable sources (search, calculators, databases) is a cleaner architecture.
- gpt-oss innovation: OpenAI's first open-weight model with native tool use training is significant — it normalizes this pattern in the open-source ecosystem.
- Trust barrier: Sebastian notes the adoption challenge: "You don't want to run this on your computer where it has access to tools and could wipe your hard drive." Sandboxing and permissions are unsolved UX problems.
- Not a complete solution: Tool use helps with factual queries but doesn't address reasoning hallucinations or plausible-sounding fabrications. It's mitigation, not cure.
Verdict: Valid architectural direction with practical limitations
📊 DeepSeek "Losing Its Crown"
"I would say that DeepSeek is kind of losing its crown as the preeminent open model maker in China, and the likes of Z.ai with their GLM models, MiniMax's models, and Kimi K2 Thinking from Moonshot... have shone more brightly."
Our Analysis
Nathan's observation about the Chinese AI ecosystem fragmenting is noteworthy:
- DeepSeek's legacy: The January 2025 DeepSeek R1 moment was transformative — it proved open-weight models could match frontier performance. That impact is permanent even if leadership shifts.
- Competition is healthy: Z.ai, MiniMax, and Moonshot having filed IPO paperwork signals commercial maturation. More players = faster iteration.
- The nuance: Sebastian pushes back that DeepSeek isn't worse — others are using DeepSeek's innovations. The "crown" metaphor may overstate the competitive shift.
- Secretive vs. promotional: DeepSeek's hedge fund backing (Highflyer Capital) means less need for marketing. Startups seeking Western investment are naturally louder.
Verdict: Possibly overstated — ecosystem diversifying rather than DeepSeek declining
What Should We Believe?
This episode is refreshingly grounded compared to typical AI hype. Lambert and Raschka are practitioners, not salespeople. Here's what to take away:
- The US-China race is real but not winner-take-all: Ideas flow freely, talent moves, and both ecosystems produce frontier models. The differentiator is compute access and operational execution, not secret breakthroughs.
- Model differentiation is narrowing: The conversation about switching between ChatGPT, Claude, and Gemini for different tasks reveals that no single model dominates. User habits matter more than marginal capability differences.
- Open-weight is strategic: Whether from China or OpenAI, releasing open models is calculated — about distribution, ecosystem influence, and in OpenAI's case, offloading compute costs. This isn't altruism.
- Compute constraints are binding: GPU scarcity shapes industry dynamics in underappreciated ways. OpenAI's "GPU deprived" comments and Google's TPU advantage are about economics, not just technology.
- The Claude Code hype is real but narrow: For developers, Claude with extended thinking may genuinely be best-in-class. For the other 99% of AI users, it's largely irrelevant to their experience.
The Bottom Line
This is the kind of AI discussion that's actually useful: technical enough to be substantive, honest about uncertainty, and free from obvious commercial agendas. Lambert and Raschka represent the researcher/engineer perspective that often gets drowned out by CEO proclamations and investor narratives.
If you work in or around AI, this episode provides excellent mental models for how the landscape actually functions. If you're a casual observer, you'll learn that the reality is messier and more competitive than the "OpenAI vs. everyone" framing suggests. The AI race is a marathon with many runners, not a sprint with predetermined winners.