Overview
8 stories in this issue. The first 3 are today's priorities.
Hot model updates
- Top · Gemini 4 Carbon tops MathArena at about 15x GPT-6.1 Sol's cost
- Top · A developer builds a 190-line pure-Jax demo of the DeepSeek V3 architecture
Global AI news
- Top · Amodei says compute must be bought one to two years ahead; Anthropic's commitments near $901B
- After Jensen Huang's exit, South Korea reviews Nvidia GPU supply; report says ministry cites unchanged 260,000-GPU plan
- AT&T runs nearly 500 agentic workflows in production and processes 45 billion tokens a day
- TechCrunch's "AI graveyard": Character.ai files for bankruptcy, and about 42% of corporate AI projects are abandoned
Regional and early signals
- An Apple job posting shows plans for an MCP-connected AI data lake in its marketing group
- Anthropic is hiring staff engineers to build "Claude-first" internal software from scratch

Jiufeng graphic based on the sources cited in this issue.
Hot model updates
01/08
Gemini 4 Carbon tops MathArena at about 15x GPT-6.1 Sol's cost
Gemini 4 Carbon pushed GPT-6.1 Sol off the top of MathArena. At the reasoning settings used in testing, though, it costs about 15 times as much as Sol.
According to The Decoder, Google's Gemini 4 Carbon (high) now leads the MathArena math benchmark. It is the first non-OpenAI model to hold the top spot in a while. Anthropic's Claude Fable 5.2 came in second, and the former leader, GPT-6.1 Sol, dropped from first place.
| Model | MathArena rank | Relative cost (GPT-6.1 Sol = 1) |
|---|---|---|
| Gemini 4 Carbon (high) | #1 | ~15x |
| Claude Fable 5.2 | #2 | ~40x |
| GPT-6.1 Sol | Former #1, displaced | 1 |
Limitations: The cost multiples reflect the reasoning settings used in testing. Carbon's most common failure is running past the time limit, and it almost never gives up on a task. In one case it launched a script that ran for more than eleven hours without checking on its progress. The report also notes that OpenAI's newer models, which reportedly solved problems like the Navier-Stokes equations, are not publicly available and were not part of the comparison.
Source: Original source Jasper Dekoninck (X) · The Decoder
02/08
A developer builds a 190-line pure-Jax demo of the DeepSeek V3 architecture
A GitHub project implements the DeepSeek V3 model architecture in 190 lines of pure Jax and reached 145 points on Hacker News.
Developer Vaibhaav Tiwari published the project on GitHub. The README and the author's comment both say 190 lines, while the Hacker News title says "~200 lines." The project was posted as a "Show HN" and reached 145 points within eight hours.
- Size: 36.2M parameters, pretrained from scratch
- Hardware: a MacBook Air
- Training data: 207 detective novels
Commenters mostly treated it as a way to learn a model from first principles. One pointed readers to jaxformer, Karpathy's nanochat and the original "Attention Is All You Need" paper.
Limitations: The README describes the project as "an architecture demo, not a strong language model." It is a personal teaching project, not an official DeepSeek release.

Image source: GitHub.
Source: Original source GitHub · Hacker News
Global AI news
03/08
Amodei says compute must be bought one to two years ahead; Anthropic's commitments near $901B
Anthropic CEO Dario Amodei called buying compute "the deepest business dilemma" his company faces. Capacity has to be committed one to two years before anyone knows whether revenue will cover it.
Speaking with Jamie Dimon at a JPMorgan Chase investor event, Amodei said: "There's a huge uncertainty in what's going to happen next year." The Decoder, citing Bloomberg, reports that Anthropic's commitments total nearly $901 billion, mostly for compute and chips. About $71 billion of that is due by September 2027.
The company plans to go public on November 19. According to the report, one goal of the listing is to build a cash buffer in case spending and revenue don't line up.
Limitations: The roughly $901 billion commitment figure is as reported by Bloomberg. Amodei described the uncertainty but gave no revenue forecast.
Source: Original source Fireside video (X) · The Decoder
04/08
After Jensen Huang's exit, South Korea reviews Nvidia GPU supply; report says ministry cites unchanged 260,000-GPU plan
After Jensen Huang resigned on November 12, South Korea moved to assess whether Nvidia's GPU supply and technical support would be affected.
RuntimeWire reports that South Korea's Presidential Council on National Artificial Intelligence Strategy held an emergency meeting on November 13 to assess how the leadership change could affect GPU deliveries and technical support for Korean AI projects. The report also says Korea's science ministry stated that Nvidia reaffirmed its 260,000-GPU supply plan.
Limitations: The full RuntimeWire article is paywalled and we could only see its opening, so we could not verify the ministry's exact wording. The 260,000 figure is the Korean government's statement as relayed by the report, not an Nvidia announcement.
Source: Single report RuntimeWire
05/08
AT&T runs nearly 500 agentic workflows in production and processes 45 billion tokens a day
AT&T has moved past using AI for employee productivity and now uses autonomous agents to redesign whole business processes. It plans to triple its number of workflows next year.
SiliconANGLE reports on AT&T's agentic AI work, citing Luke Corbin, AT&T Business's assistant vice president of digital transformation and AI, who spoke at an AWS event in Boston this month:
- Production workflows: nearly 500 agentic workflows live, with plans to triple that next year
- Token volume: Corbin says the company processes about 45 billion AI tokens a day
- Transactions: more than 1 million a month on the agentic platform
- Autonomy threshold: agents run autonomously only after accuracy reaches the mid- to high-90% range within a permissible risk profile
- Returns: Corbin says production AI investments are returning about five times their cost within a year
Limitations: These claims come from AT&T's presentation at a vendor event and so far appear only in SiliconANGLE. The report does not name the models used or give specific cost figures.
Source: Single report SiliconANGLE
06/08
TechCrunch's "AI graveyard": Character.ai files for bankruptcy, and about 42% of corporate AI projects are abandoned
TechCrunch has started a running list of AI projects that shut down or missed expectations, from startups to products at Apple and OpenAI.
The list includes chatbot platform Character.ai, which filed for bankruptcy last month just five years after launch. It also includes Ollie, a privacy-focused AI assistant that announced its shutdown a few weeks ago. Citing S&P Global Market Intelligence, the report says about 42% of AI initiatives are eventually abandoned by their parent companies. The reasons include insufficient funding, technical problems, competition, difficulty scaling and weak user demand. Apple's repeatedly delayed Siri AI and OpenAI's messy "super app" launch are also on the list.
Limitations: This is a roundup, and the source explains each project's failure in varying detail. The 42% figure is an industry statistic, not about any one company.
Source: Single report TechCrunch
Regional and early signals
07/08
An Apple job posting shows plans for an MCP-connected AI data lake in its marketing group
A role Apple posted on November 13 describes a central data lake for the AI tools, models and workflows used by Marcom, its global marketing organization.
RuntimeWire reports that the role is Senior Data Governance Manager, Marcom. According to the description, operational tools would stop keeping their own copies of Marcom data and would connect to the lake through APIs and Model Context Protocol (MCP) interfaces. Data used for training, fine-tuning and inference would go through disclosure, provenance and validation controls.
Limitations: This is a job posting, not a product announcement, and it gives no launch date.
Source: Single report RuntimeWire
08/08
Anthropic is hiring staff engineers to build "Claude-first" internal software from scratch
Anthropic is recruiting Staff Software Engineers for Internal Products to build a new set of internal software centered on Claude.
RuntimeWire reports that the posting describes shared APIs and MCP tools, scoped access, evaluations, audit trails and human confirmation points.
Limitations: The posting has no publication or update date, and Anthropic gives no deployment date. The report stresses that this is a planned new build, not a customer-facing product release.
Source: Single report RuntimeWire
This digest is compiled by AI from the sources linked in each item and published after automated fact-checking, without article-by-article human review before publication. If anything differs from the original, the original source prevails. How this digest is made, or report an error: About this digest
Create AI images and videos with sound in your browser: write a prompt, pick a model, download the result. 5 free credits on sign-up, no credit card.
Start creating free

