New AI models, token minimization and IBM’s new sub-1nm chip | 26 Jun 2026

Nuggets

New AI models, token minimization and IBM’s new sub-1nm chip | 26 Jun 2026 | Mixture of Experts · Nuggets

Sakana AI's "Fugu" Model: A highly agentic model out of Japan that has captured attention by matching or beating established open benchmarks (like LiveCodeBench) compared to prominent systems like Mythos [00:14:22].
- Architecture Reality: Fugu is not a net-new foundational model from scratch; it is an intelligent multi-agent orchestration platform [00:17:24]. It features a single novel router model that carefully channels requests behind a unified API (accepting typical JSON role/content payloads) to a network of existing frontier models [00:17:24].
- Trade-offs: The setup provides enterprise resilience against model fluctuations [00:18:24]. However, it drastically increases non-determinism, as quality can change wildly depending on how effectively the request is routed behind the scenes [00:18:42].
- Future Outlook: The team notes that the "Malaya" project out of IBM works similarly with virtual endpoints [00:22:44]. While chasing massive frontier models like Anthropic's Fable or Mythos drives headlines, the real horizontal innovation will be bringing these orchestration harnesses to a smaller scale capable of running locally on smartphones or commodity laptops [00:22:51].
ZAI's GLM 5.2: A massive open-weights coding model originating from a Chinese laboratory [00:24:24].
- Scale and Usability: The model's footprint is enormous, rivaling the size of proprietary frontier options like Claude Sonnet 4.6 [00:28:30]. Because of its extreme size, it hits "usability gates" for individual tinkerers who cannot run it locally without industrial-grade setups (like a DGX Spark) [00:24:40]. Enterprise security policies also restrict sending internal code out to it [00:25:06].
- Technical Elements: Highly inspired by the latest DeepSeek architecture, it successfully introduces speed and quality over an extremely long context length to an extremely large model [00:26:14].
- Quirks: Due to a heavily Mandarin-centric training dataset, the model occasionally blurts out random Mandarin phrases even when prompted and interacting entirely in English [].

The Corporate Context: Corporate America previously entered a phase of "Token Maxing," evaluating AI adoption by measuring total raw token consumption on the assumption that more volume directly equated to higher productivity [00:39:00]. However, enterprises have run into massive budgetary shocks, with companies like Uber and Microsoft finding their yearly cloud API budgets entirely swallowed up within 3 to 4 months [00:40:03].
The Movement to Token Minning (Minimization): Spurred in part by an article from Eli Tan at The New York Times ("Tech workers maxed out their AI use, now they're trying to minimize it"), the trend has shifted aggressively to token economics—maximizing the business value out of every single unit token [00:39:22].
Predictive Metrics of Consumption: Gabe references an adjacent article by Steve Yegge titled "The Flat Curve Society" [00:42:31]. Data shows that up to a threshold of spending about 5 million to 15 million tokens a day, raw token volume is a relatively decent predictor of overall productivity and worker output [00:43:25]. Beyond that threshold, it becomes a terrible predictor due to the gamification of metrics [00:43:40].
The True Target (Efficiency & Offloading): The panel notes that "minning" can be a misleading antonym to "maxing" because not all tokens carry equal cost. Tokens run on a local machine are practically free and can be safely maxed out [00:44:23]. True token optimization relies on smart orchestration and local offloading:
- An estimated 50% of routine technical workloads (basic internet research, repository code summaries, information digging, foundational code templates) do not require multi-million dollar cloud frontier models [00:45:20]. They can be effectively serviced by 30B to 100B parameter models running locally on standard enterprise hardware [00:45:34].
Historical Anecdote: Tim Hong shares that the premise of integrating small neural nets into household items dates back to the 1980s or 1990s, when a consumer rice cooker was sold utilizing an early old-school perceptron network solely designed to measure moisture levels to optimize rice cooking [00:38:20].

New AI models, token minimization and IBM’s new sub-1nm chip | 26 Jun 2026 | Mixture of Experts

1. Special Segment: Sub-1 Nanometer Chip Technology & "NanoStack"

References

More nuggets

Why PhonePe bet on UPI when everyone chose wallets | Sameer Nigam | Unstarted | Z47 Moments

Building a Power Company for the Next 50 Years | Zach Dell, Base Power | David Senra

What 12 Years of Failure Taught Steve Jobs About Success | 1 Sept 2026 | The Knowledge Project Podcast

New AI models, token minimization and IBM’s new sub-1nm chip | 26 Jun 2026 | Mixture of Experts

1. Special Segment: Sub-1 Nanometer Chip Technology & "NanoStack"

References

More nuggets

Why PhonePe bet on UPI when everyone chose wallets | Sameer Nigam | Unstarted | Z47 Moments

Building a Power Company for the Next 50 Years | Zach Dell, Base Power | David Senra

What 12 Years of Failure Taught Steve Jobs About Success | 1 Sept 2026 | The Knowledge Project Podcast

2. Emerging Frontier & Open-Weights Models (Sakana Fugu & GLM 5.2)

3. Entertainment & AI Industry Collaborations (Google DeepMind & A24)

4. The Shift to "Token Minning" & Efficiency Economics

EP. 20 | 10,000 Pitches, 55,000 Deals, and AI: How Pranav Pai Finds Winning Startups | 2 Sept 2026 | Clearing The BLUR