Large Language Models

Today's Stories

1. 1

   Salesforce's Agentforce hits adoption wall; only 34% of customers on board

   Salesforce CEO Marc Benioff positioned Agentforce as the company's flagship autonomous AI platform when it launched in 2024, but adoption has stalled. Only 34% of customers have adopted it, and analysts estimate just 23,000 of the company's 150,000 customers are actually using the platform. The stock has fallen more than 50% from its December 2024 peak, erasing more than $200 billion（約32兆円） in market value. The slow adoption reveals that enterprises aren't ready for agentic AI — not because they doubt its value, but because their data infrastructure isn't prepared. KeyBanc and Bernstein's research identified two core blockers: many companies struggle with fragmented CRM records and disconnected systems, and Agentforce remains in early stages, with most deployments still limited to proof-of-concept projects rather than full rollouts. For marketers, this means investing in data quality and integration will likely yield better returns than deploying AI agents on messy data.

   Salesforce is working to address these barriers by adding technology to automatically pull customer data from external sources and expanding data-management capabilities through acquisitions like Informatica. The company disputes the downgrade reports, with Benioff calling KeyBanc's assessment a "bad call" and citing internal metrics showing Agentforce is the fastest-growing product in company history. However, a CIO survey found more organizations expect to reduce Salesforce spending over the next year than increase it.

2. 2

   Dave Eggers tells OpenAI staff ChatGPT is 'silencing an entire generation'

   Author Dave Eggers, invited by Sam Altman to speak to around 200 OpenAI staffers, criticized ChatGPT's impact on education. According to the Financial Times, Eggers told staff that the tool has made teachers' lives "infinitely more difficult" and that students using it to compose will "never learn to write" and have "their voice stolen from them." Eggers is a prominent literary figure—his novel The Circle critiques the tech industry, and he has previously called AI-generated writing "pastiche nonsense." His remarks reflect growing concern among educators and writers about AI's role in undermining writing skill development and student voice, a concern that carries weight when voiced inside a major AI company.

   Altman's response and whether OpenAI addresses educator concerns raised by the criticism. Eggers' track record suggests he came prepared to challenge the company; his nonprofit work supporting writers and the arts underscores his stake in the issue.

3. 3

   Intel, Google deepen AI partnership for chip design

   Intel and Google Cloud announced an expanded partnership on Thursday, 16th, deploying Gemini Enterprise across Intel's workforce and integrating agentic AI tools into Intel's chip design process. The partnership signals Intel's commitment to embedding AI into its core engineering operations. Agentic AI tools—systems that can autonomously plan and execute tasks—represent a step beyond traditional AI assistance, potentially accelerating the chip design cycle, which is a critical competitive lever for semiconductor makers.

   The specifics of how agentic AI will be deployed in chip design workflows, and whether this partnership yields measurable improvements in design speed or quality that Intel can report in coming quarters.

4. 4

   CPU demand set to surge with agentic AI; AMD best positioned

   As AI becomes more agentic (capable of autonomous reasoning), the ratio of graphics processors to central processors in data centers is expected to shift from 8-to-1 for training down to 4-to-1 for inference and 1-to-1 for AI agents, with Nvidia predicting this could become a $200 billion（約32兆円） market in the next few years. AMD, Arm Holdings, and Intel are positioned to benefit, but AMD emerges as the strongest candidate due to its leadership in data center CPUs, its new Venice architecture with up to 256 cores designed for agentic AI, and its acquisition of memory optimization company MEXT. CPUs handle the sequential reasoning that allows AI agents to stop and think before acting—a fundamentally different workload from the raw computing power GPUs provide. AMD already leads the data center CPU market and has two massive deals with OpenAI and Meta, while also benefiting from the surging inference market with its GPUs. By contrast, Intel faces a stagnant broader computer chip business and a struggling foundry operation, and Arm risks competing against its own customers while facing manufacturing capacity constraints.

   Arm projected the data center CPU market would reach $100 billion（約16兆円） over the next five years and believed it could capture 15% market share, which would equate to $25 billion（約4兆円） in 2031 revenue with $15 billion（約2.4兆円） from CPUs. AMD's Venice architecture rollout and execution on its GPU inference strategy will be critical indicators of its ability to capture share in the agentic AI market.

5. 5

   Kimi K3 AI model raises frontier questions after strong debut

   Kimi K3, a new AI model from China, has scored 57 on the Artificial Analysis intelligence index—one point ahead of Claude Opus 4.8, two behind Sol, and three behind Fable—prompting comparisons to DeepSeek's recent market impact and talk of potential stock declines for Google, SpaceX, and Nvidia. The model's strength is prompting reassessment of existing assumptions about AI development and competition; the author notes the score may overstate capabilities but signals that independent verification over the coming days will clarify its true position relative to leading models.

   A full analysis of Kimi K3 is planned for early next week; market reactions and third-party validation of the model's actual performance against leading systems will help establish whether this represents a significant frontier shift.

6. 6

   GPT-5.6 Pro solves 30-year complexity theory problem

   OpenAI's GPT-5.6 Pro proved that a 30-year-old open problem in complexity theory—whether two ways of specifying finite closure systems define the same family—is coNP-complete. The proof was generated through ChatGPT and verified by the author, who published the work on July 18, 2026. The result settles a longstanding question that blocked progress in three connected fields: Formal Concept Analysis, database theory (functional dependencies), and the enumeration of irreducible closed sets from logical implications. It establishes that no polynomial-time algorithm can solve this equivalence test in general, unless P equals NP—a foundational impossibility that affects researchers and practitioners in logic, databases, and AI.

   The author takes full responsibility for the mathematical claims and the final manuscript, and has published the proof as a peer-reviewed paper. This marks a milestone in AI-assisted mathematical discovery, where a large language model contributed the core proof of a problem described as "widely open" at a major conference (ISAAC 2025) just months before.