Last Week in AI – Episode #231 Summary

Podcast: Last Week in AI
Date: January 21, 2026
Hosts: Andrei Karenov (A), Jeremy Harris (B)
Episode Theme:
A broad yet deep roundup of recent AI news: major tool launches (especially Anthropic’s “Claude Cowork”), billion-dollar funding rounds, emergent research on memory and scaling, shifts in chip supply and infrastructure, and the latest on open source and policy. The episode is packed with discussion about AI agent safety, new architecture proposals, business competitions, and the ongoing battle for global AI dominance.

1. Main Theme and Overview

• This episode covers two weeks’ worth of AI developments, highlighting a mix of significant product releases, large funding rounds, important open source and research efforts, and pivotal policy debates.
• Key focus: How tools like Claude Cowork and engineering advances are reshaping both day-to-day workflows and the strategic landscape for companies like Anthropic, Google, OpenAI, Nvidia, and their global competitors.

2. Tools and Apps

Anthropic’s Claude Cowork Agent

[02:28–09:45]

• Claude Cowork is a major new desktop agent from Anthropic bundling “Claude code”’s agentic capabilities into an easy-to-use app.
  • Goes beyond coding: edits videos, manages spreadsheets, sorts files, and generally acts as a digital intern/assistant.
  • “Cowork is kind of just an all purpose aid in... a new way to interact with your computer.” (B, 03:38)
• Security: Runs in a sandboxed virtual machine, not directly on your device — setting a “new gold standard” for agent security.
• Business Shift: At $100–$200/mo, Cowork is not simply "selling intelligence, [they're] selling labor." (B, 09:09) This agent is priced like hiring a part-time assistant.
• Alignment Confidence: Anthropic seems "more comfortable believing their agents are not going to go off and do things they're not supposed to." (A, 07:11)
• Key Quote:
  • “You can say, hey, sort everything by file type and date or ... automatically build Excel spreadsheets ... do a bunch of work you might have an intern do ... It's a pretty broad set of capabilities.” (B, 03:54)
• Price & Strategy: High, but justified as selling labor. B2B customers expected to drive adoption: “Anthropic is dominating in that B2B segment…much more profit per token.” (B, 21:20)

Google Gemini and Search Integration

Personal Intelligence

[09:45–12:45]

• Gemini’s new “Personal Intelligence” connects to Gmail, Photos, Search, & YouTube, letting it reason over your private data.
• Guardrails: Opt-in, can disconnect anytime, with restrictions on sensitive topics.
• Open Beta: Now out for Pro & Ultra users.
• Quote: “There may be some over personalization … but certainly Google's now shipping, which is a big shift.” (B, 12:23)

AI Overviews in Search

[12:45–16:14]

• Some AI health summaries were removed after serious errors were flagged (e.g., incorrect cancer diet advice).
• Reflective of a rapid shift: “I cannot count the number of times … I'm just asking Google a question … and you look at the AI overview and that's just standard.” (A, 14:45)
• Google’s AI Overviews now serve as a “default” for many search queries; ChatGPT’s initial search threat may be receding.

Gemini Expands in Gmail

[16:14–18:11]

• Gemini inside Gmail now offers more agentic capabilities: question answering, proofing, AI Inbox (filters/clutter management).
• “Who is the plumber that gave me a quote for the bathroom renovation last year?” as a natural query. (B, 17:38)

SlackBot AI Agent

[18:11–20:04]

• Salesforce’s new Slackbot is a “super agent”: Searches information, drafts emails, schedules, and interacts with various work apps.
• “The next step … is all of the business apps … will have agents built in.” (A, 19:19)

3. Applications and Business

Mega Funding Rounds

[20:11–24:12]

• Anthropic: $10B at $350B val (double from Sept 2025). Preparing for IPO late 2026; dominates B2B, on track for break-even 2028.
• xAI (Elon Musk): $20B at $230B val, with several sovereign wealth funds involved. Leveraging X data, Tesla data, and DoD contracts.
• Quote: “You're in the territory of sovereign wealth fundraises.” (B, 20:46)

Nvidia & Chips Supply Chain

[24:12–29:26]

• H200 Chips: China has 2M H200 chip orders, but only 700K inventory. Average chip price: $27K.
• Packaging (COAS) is the manufacturing bottleneck, not fabrication.
• “This idea of exporting advanced chips to China actually hurts American companies directly … packaging is used for both Hopper, Blackwell, and China-aimed H200.” (B, 25:39)
• Ongoing political uncertainty—US/China policies and elections could change trade rules overnight.

OpenAI & Cerebras Compute Deal

[29:26–31:38]

• OpenAI signed $10B multi-year deal with Cerebras for 750 megawatts of inference compute.
• Inference-specific hardware, aiming for resiliency against supply chain interruptions.

CoreWeave’s Liquidity Maneuvers

[31:38–34:30]

• Cloud compute provider CoreWeave is renegotiating covenants to ride out GPU delivery delays.
• Lenders appear confident in business model and recovery once hardware arrives.

LMSYS/Chatbot Arena Success

[34:30–35:54]

• LM Arena (spinout from “Chatbot Arena”, originally academic/open-source) now valued at $1.7B after rapid growth, serving $30M of annualized revenue.

4. Projects and Open Source

Key Model and Architecture Releases

Nemetron Cascade: Cascaded RL for Reasoning

[35:54–43:14]

• Proposed cascaded RL fine-tuning across discrete domains (math→code→chat) using RLHF, preventing “catastrophic forgetting.”
• "When you look at cascade RL... the model is not just mimicking tokens... it reinforces the underlying reasoning capabilities." (B, 41:22)
• Open-source model + detailed training recipe from Nvidia.

DeepSeek – Manifold Constraint Hyperconnections (MHC)

[43:14–49:53]

• Improves upon residual connection architecture with mathematically rigorous “hyperconnections” (advanced information flows).
• Extensive technical explanation; solution prevents signal blowup/fading in very deep networks via doubly stochastic matrices.

iQuest Coder v1

[49:53–54:04]

• Coding-specialized models (7B–40B).
• Notably, the “40B loop” variant processes inputs in two passes, first building a global context “key-value cache,” then generating code—an innovation to help models “think twice.”
• Highly competitive on code benchmarks for its size.

Falcon H1R7B Hybrid Model

[54:04–60:23]

• 7B parameter “hybrid” (Transformer + Mamba 2), 256K context, outperforms larger models on some reasoning tasks.
• Parallel hybrid layers (running Mamba and attention heads in parallel), Mamba enables efficient long-term memory.
• “Once again, we’re seeing this Mamba-transformer merger... second or third time this year.” (B, 58:10)

5. Research & Advancements

Deep Delta Learning

[60:23–66:47]

• Theoretical work proposing new “delta operator” to generalize residual stream connections, giving the model flexibility to “flip” or “delete” features, not just add.
• May remove a fundamental transformer limitation: “If you believe that neural nets need to be able to delete bad features ... this matters.” (B, 63:55)

Recursive Language Models

[66:47–72:59]

• MIT paper: Models treat prompts as “external environments,” recursively calling themselves to process arbitrarily long texts (books, datasets).
• Enables dynamic, programmable searches for information, building on subagents/scripts inside LLM pipelines.

Conditional Memory via Scalable Lookup (DeepSeek)

[72:59–78:15]

• Proposes adding "ngram" lookup blocks within models—augmented, learnable external memories—distinguishing stored knowledge vs. reasoning.
• Bridges the gap between memorization and generalization; potentially essential for future LLM scaling.

Extended Contexts by Dropping Position Embeddings

[78:15–85:34]

• Demonstrates models can shed position embeddings (like RoPE) after pretraining, enabling much larger context windows “for free.”
• “The model's internal layers … learn to tell time on their own … can handle massive amounts of data.” (B, 81:36)
• Fine-tuning without explicit position results in robust context generalization.

6. Policy & Safety

Anthropic: Constitutional Classifiers Robustness

[85:34–90:24]

• New approach to defeating “universal jailbreaks” (e.g., splitting/obfuscating harmful queries).
• Uses a two-stage classifier (lightweight → heavyweight) plus linear probes on model activations for efficiency.
• Considerable red teaming, with bounties up to $35K for jailbreaks detected.
• “Very, very low refusal rate … 0.05% on production traffic.” (B, 87:56)

Nvidia/China Exports & Regulatory Comments

[90:24–94:43]

• Jensen Huang: China will not make splashy chip-buy announcements; approval = purchase orders.
• Chinese AI execs public about compute shortfalls: “less than 20% probability to leapfrog OpenAI/Anthropic …”
• US export controls have real, persistent impact; inference-time compute is the new bottleneck.

US Policy: Export Controls Debated

[96:56–100:31]

• Jake Sullivan (Biden natsec advisor) slams Trump for repealing controls; Democrat “ESG”/ethics bundling blamed for partisanship.
• "There's no reason that being a Republican or Democrat should affect whether or not you think friggin chip sales should happen to China. That's a crazy thing." (B, 98:41)
• Money and big tech lobbying play critical roles regardless of party.

7. Notable Quotes & Memorable Moments

• “Cowork is just … an all purpose aid … like, a new way to interact with your computer.” (B, 03:38)
• “Anthropic is dominating in that B2B segment. That’s a much more … profit per token.” (B, 21:20)
• “Once again, we're seeing this Mamba-transformer merger …” (B, 58:10)
• “The model is not just mimicking tokens. … it reinforces the underlying reasoning capabilities.” (B, 41:22)
• “It is a theory [delta operator] … but if you believe networks need to be able to delete … this matters.” (B, 63:55)
• “Nvidia's buying [AI21 Labs], or at least in talks to buy them, which is very interesting … completely, completely missed that.” (B, 92:01)
• “There's no reason being a Republican or Democrat should affect chip sales to China … That’s a crazy thing.” (B, 98:41)
• “With the Chinese labs themselves telling us our policy should be … at least from them … seems pretty clear … the chips are a big thing, right?” (B, 94:43)

Important Timestamps

• Anthropic ‘Claude Cowork’ deep dive: [02:28–09:45]
• Gemini ‘Personal Intelligence’: [09:45–12:45]
• AI Overviews health issue: [12:45–16:14]
• Slackbot as Agent: [18:11–20:04]
• Anthropic/XAI Funding: [20:11–24:12]
• Nvidia H200 chips/China: [24:12–29:26]
• CoreWeave liquidity: [31:38–34:30]
• Open Source model deep-dives: [35:54–60:23]
• Research/Delta Operator: [60:23–66:47]
• Recursive Models: [66:47–72:59]
• Policy: US/China chips: [93:13–100:31]

Closing Notes

• The hosts stress that modern AI research is now so transparent that “there really aren’t many secrets”; with massive funding, most tech can be replicated.
• The pace and openness of new architectures, coupled with real-world problems (chips, regulation, market), will keep AI fiercely competitive and unpredictable across borders.

End of Summary.