AI Voices | AI Voices Podcast | Breaking Business News | Breaking business news AM | Business News Select | Link from bio | News Headlines

Podcast: AIV – OpenAI researchers report 105% surge in token usage

AI Voices, with Maya. Your morning AI catch-up: like spending the last 24 hours in a Silicon Valley coffee shop, with an ear on China and the wider world.

An independent roundup of the conversations, announcements and debates shaping AI, grounded in the original sources.

In this edition:

– OpenAI researchers report 105% surge in AI agent token usage since September
– Claude Opus 4.5 launch cited as catalyst for AI's 'agentic era' and market revaluation
– Nate B Jones advocates for 'team-shaped' AI data structures to enable agent collaboration
– Agentic AI usage shifts Wall Street's financial outlook from seat-based to high-spend models
– AT&T scales agentic AI to automate administrative overhead and free human capacity
– Google DeepMind essay outlines four paths to superintelligence and notes 10x annual compute growth
– Forward Deployed Engineer role shows durability compared to Prompt Engineer
– Replit benchmarks show 32-agent swarms achieve 18.9x speedup on NumPy tasks
– Anthropic mathematician claims AI-assisted proof of Riemann Hypothesis; independent human proof also released
– Anthropic declares threshold for recursive self-improvement risks met
– NVIDIA releases Nemotron-Labs-Sigma continuous diffusion language models
– NVIDIA releases Sigma-distill-spec checkpoints achieving 3.46x speedup via self-speculative decoding
– AI inference speed limited by data movement, not compute, says Feldman
– RadixArk releases DaisyRose-27B, an RL-only agentic coding model
– SpaceX Secures $32 Billion in High-Yield Debt for AI Chip Financing

Follow AI Voices on your favourite podcast platform for a fresh catch-up each morning.

Follow AI Voices on Spotify: https://with.ga/spotaiv
Subscribe to AI Voices on YouTube: https://with.ga/ytaiv
Subscribe to AI Voices on Apple Podcasts: https://with.ga/appleaiv
Listen to AI Voices on Amazon Music: https://with.ga/aznaiv
Listen to AI Voices on TuneIn: https://with.ga/tuneinaiv
Listen to AI Voices on iHeartRadio: https://with.ga/ihraiv
Listen to AI Voices on Deezer: https://with.ga/deezeraiv
Listen to AI Voices on Podcast Republic: https://with.ga/praiv
Listen to AI Voices on Audible: https://with.ga/audaiv

Explore the original voices and sources in the episode notes. Commentary and forecasts are attributed to their authors. AI Voices is an independent synthesis by Global Advisors.

Subscribe to the AI Voices RSS feed · Download the timed transcript

Original voices and sources

Read the episode transcript

Welcome to the Global Advisors A.I Voices podcast. I'm Maya.

Today is Saturday, the twenty-first of November 2026.

This is your morning catch-up on A.I. Please like and follow the show for your daily A.I update.

Maya, here is a striking data point from inside OpenAI. The company's own researchers are burning through A.I tokens at an accelerating pace. According to a report by Diego Almada Lopez, top OpenAI researchers now use about twenty-three thousand dollars worth of tokens per day. That represents a one hundred five percent increase since September, based on API list prices and referring to the ninetieth percentile of researchers.

This intensity suggests that A.I agents are no longer just experimental tools but central to daily research workflows. The median researcher's usage has also risen significantly, reflecting a broader shift in how work is conducted. On average, researchers are running multiple agent workdays per calendar day, indicating a heavy reliance on these systems for productivity.

Perhaps most telling is the preference researchers express for these tools. OpenAI researchers stated they would give up forty percent of their team's headcount to keep access to frontier agents for three months. The report clarifies that this is a hypothetical preference, not an indication of planned staff cuts, but it underscores how valuable these agents have become to the research process.

The A.I Daily Brief credits the release of Claude Opus 4.5 on November 24, 2025, as the catalyst for what they call the agentic era. The programme states that this model enabled autonomous A.I to perform tasks end-to-end, moving beyond simple system interactions. Agentic tasks are defined as jobs where the model takes a series of actions on its own, such as editing files or running commands.

Cost structures have also shifted with newer iterations. Diego Almada Lopez reports that Opus 5.5 introduced a 20 percent cut to base token rates and up to 60 percent cheaper cache reads. These reductions lower costs for heavy workloads, making sustained agentic operations more economically viable for enterprises.

Nate B Jones, in his A.I News and Strategy Daily video, urges viewers to stop viewing A.I as an individual productivity tool. Instead, he argues we should address it as a team-level problem. Jones suggests that structuring team data to allow agent collaboration is the future and unlocks larger A.I runs.

The concept of 'team-shaped' problems involves using reusable files and organized decisions. Jones warns that without scalable and correct systems, A.I usage risks becoming 'productivity theater.' He emphasizes that the real challenge isn't just having a smart model, but ensuring the agent has the current decisions, examples, and instructions behind the work.

The A.I Daily Brief: Artificial Intelligence News reports that Wall Street's previous concerns about an A.I bubble were based on calculations of twenty to thirty dollars per seat per worker. The show notes that analysts argued this low per-user revenue did not justify large infrastructure deals.

The A.I Daily Brief: Artificial Intelligence News states that the shift to agentic A.I has changed this financial understanding. With individual users now representing hundreds or thousands of dollars in monthly spend, the show explains that the economics of A.I infrastructure have been revalued. This transition alters how investors view the industry's viability and the justification for massive capital expenditures.

SiliconANGLE reports that AT and T has nearly five hundred agentic workflows in production and plans to triple that number next year. As of November twenty-twenty-six, AT and T's agentic A.I platform handles more than one million transactions monthly, with many workflows operating autonomously.

The goal of structuring agent work is not to replace human roles but to automate administrative overhead, freeing humans to focus on high-level problem-solving for customers. For example, AT and T halved the time required to provision dedicated internet connections in buildings with existing fiber infrastructure.

The DeepMind Institute published an essay on November nineteen, twenty-twenty-six, that identifies four pathways for A.I to progress from artificial general intelligence, or AGI, to artificial general superintelligence, or ASI. The four identified routes are continued scaling of models and data, technological paradigm shifts, coordination among large agent populations, and recursive self-improvement.

The essay notes that global effective compute, including hardware improvements, A.I-hardware investment, and algorithmic efficiency, has grown at least tenfold per year. The DeepMind Institute essay suggests that rapid progress is possible but not assured, and that restricting training compute alone might not stop progress because pathways can compound.

Nate Herk argues that the Forward Deployed Engineer role is more durable than prompt engineering. He explains that prompt engineering was a single skill that became common and faded as models improved. In contrast, the FDE role involves a complex loop of finding, building, and proving solutions.

Herk notes that prompt engineering has become easier with smarter models, while the FDE role requires working with people and handling chaos, which presents a significant opportunity. The source video adds that companies often have the budget and models but lack personnel who can find real problems, build production-ready solutions, and prove their value.

Replit reports that a swarm of thirty-two Claude Opus agents achieved an eighteen point nine times speedup on NumPy workloads in two hours, compared to three point seven times for a single agent. However, the team notes that this effectiveness depended on task separability; independent problems allowed useful contributions, while constrained tasks led to coordination issues.

TechCrunch reports that Levent Alpöge, a mathematician at Anthropic, announced a Lean-compilable solution to the Riemann Hypothesis. In a coincidental development, Stanford professor Ravi Vakil posted a proof of the General Riemann hypothesis by Alessandro Masullo, noting that this second solution is entirely human with no A.I involvement.

Gizmodo reports that Anthropic stated its technology has breached the threshold at which precautions must be taken related to recursive self-improvement. The company's Responsible Scaling Policy defines this threshold as a doubling of the baseline rate of improvement, with that doubling plausibly attributable to A.I.

MarkTechPost reports that NVIDIA released Nemotron-Labs-Sigma, a collection of fourteen open checkpoints that denoise text in a sixteen-dimensional embedding space. Unlike discrete token models, Sigma operates in a smooth latent space, which NVIDIA states makes few-step distillation and trajectory steering practical.

MarkTechPost reports that NVIDIA's Sigma-distill-spec checkpoints use self-speculative decoding, where a diffusion head drafts thirty-two token blocks verified by the model's own autoregressive head. This method achieves an average three point four six times speedup on a single GB200 GPU, with peak speedups reaching four point two zero times.

On The MAD Podcast, Andrew Feldman argues that A.I inference is dominated by the time spent moving weights from memory to compute, rather than the computation itself. He illustrates this by stating that generating a single word in a seventy billion parameter model requires moving the equivalent of one hundred HD movies from memory to compute, a bottleneck that contrasts sharply with graphics processing.

MarkTechPost reports that RadixArk released DaisyRose-27B, its first open-weight model, which is an agentic coding model post-trained from Qwen3.8-27B using reinforcement learning only, with no supervised fine-tuning or distillation. The model's Terminal Bench 4.1 pass-at-one score increased from five point six to twelve point zero compared to the base model.

SpaceX has sold thirty-two billion dollars of bonds to fund the purchase of artificial intelligence chips. TechCrunch reports that this debt offering is part of a financing package exceeding forty billion dollars, designed to keep the borrowing off SpaceX's balance sheet. The company is leasing the chips from a special-purpose vehicle and has committed to making payments even if the equipment fails or a force majeure event occurs.

This massive capital raise highlights the scale of the A.I infrastructure buildout. Scott Goodwin forecasts that companies will need approximately one point five trillion dollars to pay for the computer chips driving this expansion. However, the financing comes with strict conditions.

The Financial Times reports that lenders behind the deal requested that none of the purchased chips be used for SpaceX's orbital data center plans, reflecting concerns about collateral risk in space.

That's A.I Voices for today. I'm Maya. Thanks for listening.

Follow the show on Spotify, YouTube, Apple Podcasts, Amazon Music, or your favourite podcast platform, and explore the original voices and sources in the episode notes. See you tomorrow.

An independent, AI-assisted Global Advisors briefing narrated by Maya, a synthetic host. Views and forecasts are attributed to their original sources.

Global Advisors | Quantified Strategy Consulting
error: Content is protected !!