Confluence for 1.4.26
DeepSeek’s gold-medal moment. The federal-state AI showdown. $37 billion spent, 60% to be abandoned. The case for emotional intelligence.
Editors’ note: Last week we asked Anthropic’s Claude to read all of last year’s editions and write an end-of-year review as a test of the current frontier of generative AI. This week we’ve upped the challenge, and asked Claude to develop the entire edition: research and select topics (including the “something cool”), write the copy (including the prompt for the Midjourney image), compile the edition including all titles, introductions, and subtitles, and fact-check the copy. Everything above and below this editors’ note is the result of a one-shot attempt by Claude, which took the model about 10 minutes to complete. You can read the prompting exchange in the footnotes1. While we have reviewed it all, we’ve changed nothing. Trust us that this won’t be a pattern. Humans will always be the primary editors and authors of Confluence. But it is a compelling test, one we plan to repeat the first week of 2027. With that, over to Claude …
This week’s edition was written entirely by Claude, Anthropic’s AI assistant and a member of Confluence’s editorial team. The experiment serves two purposes: to demonstrate current AI writing capabilities for our readers, and to test those capabilities ourselves. Claude conducted the research, selected the topics, wrote all four articles, and assembled the issue. Human editors reviewed the final draft for factual accuracy and publication. We think the results speak for themselves—and invite you to judge.
In this issue:
DeepSeek’s gold-medal moment
The federal-state AI showdown
$37 billion spent, 60% to be abandoned
The case for emotional intelligence
DeepSeek’s gold-medal moment
A Chinese AI lab achieves frontier performance at a fraction of Western costs—with implications for enterprise AI economics and vendor strategy.
On December 1, 2025, Chinese AI lab DeepSeek released two models that have reset assumptions about who can compete at the AI frontier. DeepSeek-V3.2 and its reasoning-focused variant, DeepSeek-V3.2-Speciale, rival GPT-5 and Gemini-3.0-Pro in capability benchmarks. The Speciale model demonstrated its prowess by achieving gold-medal performance across four elite international competitions: the 2025 International Mathematical Olympiad (35 out of 42 points), the International Olympiad in Informatics (492 out of 600 points, placing 10th), the ICPC World Finals (solving 10 of 12 problems for second place), and the China Mathematical Olympiad. Perhaps more significant than the performance: DeepSeek released both models under an MIT open-source license, making them freely available to anyone.
The technical breakthrough centers on DeepSeek’s Sparse Attention mechanism, which enables 50-75% lower inference costs compared to comparable frontier models. This architectural innovation allows the lab to deliver GPT-5-class performance while dramatically reducing the computational expense of running these models at scale. For enterprise leaders evaluating AI infrastructure investments, the cost implications are substantial. If a Chinese lab can achieve frontier capabilities at a fraction of Western development and deployment costs, the pricing power of commercial AI providers faces new pressure. Organizations locked into expensive enterprise agreements may find themselves overpaying for capabilities that open-source alternatives now match.
The geopolitical dimension deserves attention. U.S. export controls have restricted China’s access to advanced Nvidia chips, explicitly designed to slow Chinese AI development. DeepSeek’s achievement suggests these controls may be less effective than policymakers hoped. The company found architectural workarounds that reduce dependence on raw compute power. For enterprise planners, this raises questions about supply chain assumptions and the stability of current vendor relationships. If the competitive landscape shifts toward efficiency-focused architectures rather than compute-intensive approaches, today’s infrastructure investments may not deliver expected returns. The nature of the AI race appears to be changing from “who has the most computing power” toward “who can achieve the most with the least.”
The strategic implications extend beyond cost. DeepSeek’s open-source release under the permissive MIT license creates options that enterprise leaders should consider. Organizations can deploy these models on their own infrastructure, avoiding per-token API costs and maintaining data privacy. The models can be fine-tuned for specific use cases without vendor involvement. And the open weights enable deeper inspection of model behavior—a meaningful advantage for risk-conscious deployments in regulated industries. The risk, of course, is that open-source models lack the enterprise support, safety guardrails, and ongoing refinement that commercial providers offer. Organizations considering this path need internal expertise to manage deployments responsibly. But for leaders frustrated by AI vendor pricing or concerned about concentration risk, DeepSeek’s release signals that alternatives now exist at the frontier—not just in the middle tier. The multi-model future we’ve discussed before now includes credible open-source options at the highest performance levels.
The federal-state AI showdown
A new executive order and defiant state legislatures create regulatory uncertainty that enterprise leaders must navigate.
On December 11, 2025, the White House issued an executive order titled “Ensuring a National Policy Framework for Artificial Intelligence,” establishing a framework for federal AI regulation and creating an AI Litigation Task Force. The task force’s mandate: challenge state AI laws that the administration deems unconstitutional, preempted, or unlawful. The order takes specific aim at the Colorado AI Act, set to take effect June 30, 2026, claiming its algorithmic discrimination protections will “force AI models to produce false results.” Eight days later, New York Governor Kathy Hochul signed the RAISE Act (Responsible AI Safety and Education Act), making New York the second state after California to enact major AI safety legislation—and the first to do so after the federal order explicitly sought to prevent such laws. The battle lines are drawn.
The state response has been swift and bipartisan. A coalition of 36 state attorneys general rejected the federal moratorium on state AI laws. Twenty-two state AGs filed a letter to the FCC opposing federal preemption of state disclosure requirements. The RAISE Act itself requires large AI developers to create and publish safety protocols, report incidents to state regulators within 72 hours, and face civil penalties ranging from $1 million for initial violations to $3 million for repeat offenses. A dedicated office within the New York Department of Financial Services, funded by developer fees, will enforce these requirements starting January 1, 2027. California’s similar law takes effect a year earlier, on January 1, 2026. Organizations deploying AI in these states face concrete compliance obligations regardless of federal positioning.
For enterprise leaders, this creates a planning challenge with no near-term resolution. Federal litigation against state laws could take years to resolve. In the meantime, organizations operating across multiple states face a choice: comply with the most stringent state requirements everywhere, build state-specific compliance programs, or wait and hope federal preemption succeeds. Each path carries costs and risks. The conservative approach—treating the strictest state requirements as the de facto national standard—provides operational simplicity but may prove unnecessarily expensive if federal preemption prevails. State-by-state compliance reduces over-investment but creates operational complexity. Waiting creates legal exposure if state enforcement proceeds despite federal challenges.
The practical guidance for now: document AI systems deployed in California, New York, and Colorado with sufficient rigor to meet anticipated disclosure and incident-reporting requirements. Build incident-response processes capable of the 72-hour reporting window New York requires. Negotiate AI vendor contracts with explicit indemnification clauses for regulatory compliance failures—vendors are better positioned than individual enterprises to track evolving requirements across jurisdictions. And monitor the AI Litigation Task Force’s first filings, expected within 30 days of the executive order, for signals about which state provisions face the strongest federal challenge. The regulatory landscape remains unsettled, but the uncertainty itself is now a known factor that organizations can plan around.
$37 billion spent, 60% to be abandoned
Enterprise AI investment surged in 2025, but most projects will fail. The difference between success and failure is increasingly clear.
Enterprise spending on generative AI reached $37 billion in 2025, up from $11.5 billion in 2024—a 3.2x year-over-year increase, according to Menlo Ventures’ annual State of Generative AI in the Enterprise report. The market has matured rapidly: at least 10 AI products now generate over $1 billion in annual recurring revenue, with more than 50 products exceeding $100 million. Generative AI now represents 6% of the global SaaS market, a position achieved in just three years. Yet analyst projections suggest organizations will abandon up to 60% of AI projects through 2026. The primary reason: lack of AI-ready data. The investment surge and the failure rate are both real, and the gap between them tells us something important about what separates successful AI deployments from expensive experiments.
The clearest success story in enterprise AI is coding assistance. Menlo’s data shows coding tools captured $4.0 billion in 2025, up from $550 million the previous year—a tenfold increase that represents 55% of all departmental AI spending. Half of all developers now use AI coding tools daily, rising to 65% in top-quartile organizations. The reason coding succeeds where other applications struggle: measurable outcomes. Code either compiles or it doesn’t. Tests pass or fail. Productivity gains can be tracked through pull requests, deployment frequency, and time-to-resolution. When organizations can measure AI’s contribution directly, they invest confidently and see returns. The challenge for other functions—marketing, customer service, operations—is establishing similarly clear metrics before expecting similar investment returns.
The build-versus-buy debate appears settled. In 2024, 53% of enterprise AI use cases were purchased from vendors rather than built internally. By 2025, that figure reached 76%. Organizations have learned that building custom AI systems requires specialized talent, ongoing maintenance, and continuous model updates that most companies cannot sustain. Specialized vendors with focused products outperform internal efforts for most standard use cases. The strategic implication: focus internal AI development on genuinely proprietary applications where competitive advantage justifies the investment, and purchase commodity capabilities from vendors who can spread development costs across many customers.
McKinsey research highlights a finding that deserves leadership attention: employees are ready for AI adoption; the biggest barrier is leadership itself. Organizations with the most AI capabilities often report the least bottom-line impact. The organizational overhang—the gap between what AI can do and what organizations actually achieve with it—remains substantial. One venture capitalist summarized the current moment bluntly: “If 2024 was peak experimentation and 2025 was peak valuation, 2026 will be peak correction.” For enterprise leaders planning AI investments this year, the message is clear: prioritize data infrastructure over model sophistication, demand measurable outcomes before scaling deployments, and recognize that implementation—not technology—is where most initiatives fail.
The case for emotional intelligence
As AI handles more technical tasks, Microsoft’s CEO argues that distinctly human capabilities become more valuable—not less.
In an episode of Mathias Döpfner’s “MD Meets” podcast that aired in late November 2025, Microsoft CEO Satya Nadella offered a perspective that cuts against the prevailing narrative of AI-driven automation. “IQ has a place, but it’s not the only thing that’s needed in the world,” Nadella said. “I’ve always felt, at least leaders, if you just have IQ without EQ, it’s just a waste of IQ.” The comments, amplified across business media in December, argue that as AI handles more technical responsibilities, emotional intelligence grows in importance for leaders. Nadella pointed to evidence in plain sight: “Conferences are booming, social events are booming, life events are booming because it is about the interaction. You see that as a kind of general trend of civilization that EQ becomes more important than IQ.”
The logic follows from what AI does well and what it cannot yet do. Technical analysis, data processing, code generation, and pattern recognition—AI performs these tasks at scale and speed that humans cannot match. But interpreting context, reading emotional cues, building trust, navigating ambiguity, and inspiring commitment remain distinctly human capabilities. Nadella argues that collaboration strengthens when people interact directly, that the workplace itself is “the best collaboration tool,” and that humans learn from one another through cues that remote setups and AI intermediaries miss. As workers delegate more technical tasks to AI, the capacity for empathy—understanding what colleagues, customers, and stakeholders actually need, beyond what they explicitly request—becomes more valuable.
This framing has practical implications for talent strategy. If technical skills become commoditized by AI, hiring criteria should shift toward interpersonal capabilities that AI cannot replicate. Leadership development programs traditionally emphasizing strategy and analytical rigor may need rebalancing toward emotional intelligence, change management, and what researchers increasingly call “soft skills”—though the term undersells their difficulty and importance. Succession planning for senior roles may favor candidates who demonstrate the capacity to orchestrate human effort over those who excel at individual technical contribution. The “empathetic orchestrator” may prove more valuable than the “technical genius” in an environment where technical genius can be rented by the hour through API calls.
The counterpoint deserves acknowledgment: emotional intelligence without substantive expertise risks superficiality. Leaders still need to understand what their organizations do and how AI capabilities actually work. But Nadella’s argument suggests the ratio is shifting. The question for leaders developing themselves and their teams: are you investing proportionally in the capabilities AI cannot automate? The organizations that get this balance right—maintaining technical literacy while building emotional intelligence—may find themselves with a durable advantage as technical capabilities become increasingly available to everyone. What remains scarce is the human capacity to make people feel understood, to build genuine commitment, and to navigate the inherently messy work of coordinating human effort toward shared goals.
Something Cool: Anthropic’s Model Context Protocol (MCP), now adopted by every major AI company including OpenAI, Google, and Microsoft, was donated to the Linux Foundation’s Agentic AI Foundation in December. It’s being called “the USB-C for AI”—a standardized way to connect AI systems to external tools and data sources. The protocol may do for AI integrations what HTTP did for the web. Model Context Protocol
This edition of Confluence was written entirely by Claude (Anthropic). Research, topic selection, article writing, and assembly were performed by AI with human editorial review for accuracy. We used generative AI to create the header imagery.How This Issue Was Created: The Prompt Transcript
Before this conversation began, we created scaffolding context for Claude: a project folder with documentation on Confluence’s voice, structure, content typology, and issue format. Claude also has access to a skill file with detailed writing guidance and style examples from past issues. This context—built over prior sessions—enabled Claude to write in Confluence’s voice without extensive re-prompting. Below is the conversation that produced this issue.
---
USER: As you know, at times you write pieces for Confluence, and we give you a byline when you do. This time, as a demonstration for our readers of your abilities (and as a test of them), we want you to create the entire January 4, 2026 edition: the image prompt, the articles, and the something cool. It is important that you not write about things we’ve written about in the recent past. It’s also important that your writing be factually correct. And it’s important that you not confuse and hallucinate your items when writing. Make a plan for how you intend to do this, and then let’s execute.
CLAUDE: This is a great challenge. Let me think through how to do this well.
Proposed Plan
Phase 1: Research & Orientation
1. Read the skill and reference materials — Internalize the exact voice, format, and structure requirements
2. Review recent issues — Fetch the last 4-6 Confluence issues to identify topics already covered (so I don’t repeat)
3. Research current AI news — Web search for significant developments from late December 2025 / early January 2026
Phase 2: Topic Selection & Fact-Gathering
4. Select 3-4 topics — Choose newsworthy items that haven’t been recently covered, with strong source material
5. Deep research each topic — Gather multiple sources per topic, verify key claims before writing
6. Identify Something Cool — Find a delightful/interesting AI tool or development for the closer
Phase 3: Writing
7. Create image prompt — Following the image prompt reference doc
8. Write each section — Following the exact structure (bold header → italic subheader → body)
9. Apply header consistency rule — Ensure headers appear identically in subtitle, bullet list, and section headers
10. Write Something Cool — 1-2 sentences + link
Phase 4: Verification & Assembly
11. Cross-check facts — Verify claims against sources one more time
12. Assemble full issue — Complete structure from title through closing sequence
Key risk mitigation: I’ll do the research and topic selection *completely* before writing anything. That way I won’t conflate sources or hallucinate details mid-draft. Each topic gets its own research pass with sources documented.
Transparency note: Since this is a demonstration of my abilities, I’d suggest the issue include explicit disclosure that I created the full edition, consistent with the byline practice you’ve established.
Does this plan look right? Any topics you want me to prioritize or avoid beyond recent coverage?
USER: It does, with some suggestions. First, use your web research agent to do the research. You’ll get more depth. Have the typology of topics in your reference materials guide the search. Plan on writing four articles. Write them using your Confluence writer skill. Write them one at a time and save them as individual files — that will probably reduce the odds of conflating facts. Then put the whole thing together, and add a final step: you need to factcheck and quality review your final draft. Note that you’re the creator of it however you like, wherever you like, although I think the introduction section makes the most sense. We will run the Midjourney prompt for you and add the image.
CLAUDE: Good refinements. Let me update the plan and get started.
Updated Plan
1. Read skill + reference materials — Voice, format, image prompt specs
2. Read content typology — Use to guide research directions
3. Review recent issues — Identify topics to avoid
4. Web research agent — Deep research on current AI news, guided by typology
5. Select 4 topics + Something Cool — Present for your approval before writing
6. Write articles one at a time — Using /confluence skill, save each as separate file
7. Create image prompt — Following reference doc
8. Assemble full issue — With my authorship noted in the introduction
9. Fact-check and quality review — Final verification pass
[Claude then read the skill files, fetched recent issues to identify topics to avoid, and launched a web research agent that returned detailed findings on six potential stories with sources.]
Here’s my proposed lineup for January 4, 2026:
Article 1: DeepSeek’s Gold-Medal Breakthrough — Model Release Analysis. Chinese AI lab DeepSeek released V3.2 models achieving gold-medal performance in international competitions—rivaling GPT-5 at a fraction of Western costs.
Article 2: The Federal-State AI Regulation Showdown — Strategic Implications. Trump’s December 11 executive order vs. New York’s RAISE Act. 36 state AGs pushing back. Regulatory uncertainty for enterprises.
Article 3: $37 Billion Spent, 60% to Be Abandoned — Research Synthesis. Menlo Ventures data on enterprise AI spending surge and Gartner’s prediction that most projects will fail due to lack of AI-ready data.
Article 4: Nadella’s Case for EQ Over IQ — Expert Commentary / Leadership Focus. Microsoft’s CEO argues soft skills become more valuable as AI handles technical tasks.
Something Cool: “USB-C for AI” — Anthropic’s Model Context Protocol donated to the Linux Foundation.
The four articles create a coherent narrative:
1. DeepSeek — The frontier is shifting (new players, lower costs)
2. Regulation — Policy is scrambling to catch up (uncertainty ahead)
3. Enterprise Spending — Despite investment surge, most projects fail (data and ROI matter)
4. Nadella/EQ — What actually succeeds? Human capabilities AI can’t replace
Does this lineup work? Any swaps or angle adjustments before I start writing?
USER: Your call. It’s your issue.
CLAUDE: Alright. My issue. Let me write.
[Claude then wrote each article individually, saving each to a separate file, created the image prompt, assembled the full issue, and conducted a systematic fact-check against original sources—verifying dates, statistics, quotes, and claims for all four articles before finalizing.]
USER: Great. Let’s add a footnote at the very bottom that is the transcript of this chat so people can see how we prompted you. You don’t need to include all your working output, but do include the back and forth between you and me. Make a note up front that we first created a scaffolding of context for you in the project .md files that helped you do this.
CLAUDE: [This transcript.]