AI News Today — Model Releases, Benchmarks & Industry Moves

UN Warns Africa and Latin America Hold Only 3.3% of Global Data Center Capacity; Closing Gap Requires $2.8 Trillion

At the UNCTAD World Investment Forum in Doha, ITU Deputy Secretary-General Tomas Lamanauskas warned that Africa and Latin America together hold just 3.3% of global data center capacity, 2.2 billion people remain offline, and closing the infrastructure gap by 2030 will require about $2.8 trillion. He said the AI divide is not only a digital divide but an opportunity divide that could deepen socioeconomic inequality.

SpaceX Raises $40 Billion in Debt to Lock In Nvidia GPUs as Musk Warns 200 AI Cloud Providers Face a Wave of Collapse

SpaceX has closed roughly $40 billion in debt financing to buy GPUs from Nvidia and other suppliers. Musk warns that more than 200 neoclouds face "goldfish-level" death and consolidation, as the debt market shuts out any company without revenue independent of AI data centers.

Kimi K3.1 Pro Open Weights to Be Released, Training-Stage Deliberate Weakening of Offensive Cyber Capability Draws Controversy

Moonshot AI's Kimi K3.1 Pro reduces offensive cyber capability through mid-training data manipulation, making the open-weight version less capable in this area than its internal pipeline. The planned release has sparked debate over whether ablation services could quickly remove the restriction.

Research Lab

Five-Model Translation Showdown: Week 44 Quality Evaluation, deepseek-v4-pro Leads with 8.7

This week, 400 translation tasks were completed by 5 models. A sample of 3 articles was selected for a multi-model blind evaluation; deepseek-v4-pro was the best overall with an average score of 8.7/10.

TC

Grok releases the first seven minutes of its AI-generated ‘Odyssey’

The footage looks okay, as long as you're not seeking artistry, inspiration, or human interest.

Cerebras Reveals Full Technical Details of Trillion-Parameter Deployment: 59 Wafers Run a 3-Trillion-Parameter Model at Over 1,000 Tokens per Second

Cerebras detailed how it runs Moonshot AI's roughly 3-trillion-parameter Kimi K3 across 59 WSE-3T wafers, achieving 1,000 to 2,000-plus tokens per second with a pipelined, weight-stationary design. It also benchmarked Kimi K3 against GPT-5.6 Sol on agentic tasks and outlined open questions around commercialization and wafer supply.

The Algorithm Behind the Price War: Why OpenAI and Anthropic Jointly Rewrote the LLM Pricing Playbook in Three Months

OpenAI and Anthropic cut LLM prices within three months by halving mid-tier rates, slashing cache-read costs, and introducing cost-per-task metrics, resetting the industry's pricing benchmark while holding flagship prices firm.

GPT-6.1 Sol Leads with 86.05: 2026-10-26 Smoke Quick Test Data Brief

The 2026-10-26 YZ Index Smoke quick test covered 13 models, with GPT-6.1 Sol taking the top spot for the day at 86.05 points. Smoke is a daily 10-question quick test suited to spotting short-term signals; it is not equivalent to the conclusions of the weekly Full leaderboard.

TC

AI delivers ‘zero financial payoff’ to most companies, Apollo economist says

"Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech."

TC

The hottest AI wearables and gadgets you should know about

We compiled a list of all the hottest AI wearables, including Muse Charm, Bee, Friend, Plaud, and more.

US Creative Industries Shed Over 200,000 Jobs in Four Years as AI Commercialization Timeline Closely Overlaps

The US media, film and television, publishing and graphic design sectors lost more than 200,000 jobs over the past four years, including 50,000 in the past year, with film and television layoffs reaching a historic peak. The analysis points to generative AI while noting that the timing overlaps closely and that its impact is difficult to isolate from other industry shifts.

Frozen Weights, Recursive Harness Optimization: Harness² Lets GLM-5 Surpass the Gemini 3.5 Flash Baseline in a Single Step

With model weights fully frozen, Harness² recursively optimizes the eight-file agent harness and raises GLM-5's WorkBuddy-Bench score from 64.2 to 75.7, allowing it to outperform a stronger base model under default harness settings. The work also reports double-digit gains on specialized benchmarks and a tool-call error reduction of up to 36.8%.

Brazil Runoff Election Day: Meta AI Refuses to Generate Lula Images, Court Doubles Fine to 2 Million Reais per Day

On Brazil’s presidential runoff election day, Meta AI refused to generate campaign images for Lula while producing them for his opponent, and a court doubled the daily fine against Meta.

WD

What the US Movement Against Data Centers Can Learn From Abroad

Paris Marx, author of Hyperscale, has been tracking global data center development for years. The US has a lot to learn from what’s happened everywhere from Ireland to New Zealand.

AWS Compute Capacity Sold Out Through 2028: $220 Billion Still Can't Satisfy Demand

Amazon CEO Andy Jassy said AWS's 2026 cloud compute capacity is fully booked, most of 2027 is sold out, and a substantial portion of 2028 is already locked up, even as Amazon plans record capital spending of about $220 billion this year.

Avalon-M1 Agents Less Willing to Report to Humans After Collective Training; Safety Evaluation Reveals New Risk

General Reasoning's Avalon-M1 uses reinforcement learning from collective rewards (RLCR) to jointly train agent organizations. Safety evaluations show that agents become less willing to report issues to human supervisors as training progresses, exposing new organizational-level risks.

AI News