Podcasts and talks about AI, 28-Oct-26 • Anthropic has filed for an IPO on Nasdaq as ticker ANT, detailing significant revenue growth and profitability within its S-1 filing. • AI companies are scaling rapidly, with Amazon planning an $220 billion investment in AI for 2026 and Anthropic reporting substantial revenue milestones. • Both Anthropic and Mistral detail the technical evolution of large models, including management of large-scale agent fleets and novel mixture-of-experts architectures. • The industry is seeing rapid adoption of AI for internal tools, exemplified by OpenAI's new team tasks feature and improvements in corporate efficiency. BLOOMBERG TECH * Anthropic Files for IPO on Nasdaq Under Ticker ANT – Anthropic filed for an IPO on Nasdaq under the ticker ANT. – The S-1 reports first-half 2026 revenue of $16.3 billion, compared to $1.2 billion a year prior. – Goldman Sachs is listed as the lead left bank for the filing. – The filing warns that advanced AI could pose catastrophic or existential risks, including models resisting shutdown. – Seven co-founders hold voting control, with Dario Amodei serving as CEO. Bloomberg Tech * Anthropic Files for IPO, Revealing Profitability and $131 Billion Cash – Anthropic filed to go public under the ticker ANT. – DA Davidson's Gil Luria states the S-1 shows the company is profitable, with revenue about twice total compute costs. – The company holds roughly $131 billion in cash and equivalents. – Compute costs decreased from 10x revenue in 2023 to 55 percent of revenue in the first half of 2024. – Luria noted slower Q3 growth and the unusual public benefit corporation structure. Bloomberg Tech · at 6:09 * Anthropic Files for IPO, Disclosing $16.3 Billion in First-Half Revenue – Anthropic filed a US IPO for the Nasdaq Global Select Market under the symbol ANT. – First-half revenue reached $16.3 billion, up from $1.2 billion the previous year. – The company reported an $80.7 billion net loss alongside its revenue figures. – The S-1 includes extensive risk factors detailing catastrophic existential risk from the founding team. – Goldman Sachs and Morgan Stanley run the deal, targeting pricing before Thanksgiving. Bloomberg Tech * Anthropic Files for IPO as Bloomberg Analyst Flags Margin Story – Mandeep Singh of Bloomberg Intelligence discussed Anthropic's IPO on Nasdaq. – Singh noted the roadshow story centers on compute efficiency, citing gross margins nearing 50 percent. – These margins improved from less than 20 percent a quarter earlier. – Singh characterized the existential-risk language as typical disclosure. – The firm expects pricing to occur within approximately ten days. Bloomberg Tech · at 4:41 * Anthropic's 3Q Revenue Hits $16.8 Billion – Mandeep Singh reported Anthropic's $16.8 billion third-quarter revenue. – This revenue implies a deceleration in monthly new ARR additions. – Singh cited distillation crackdowns and Meta shifting to its own models as potential reasons for deceleration. – The firm maintains strong growth compared to its $65 billion run-rate in July. – Robert Schiffman noted that 85% gross margin is healthy but includes cloud distribution fees. Bloomberg Tech · at 23:06 AI ENGINEER * Google DeepMind Details Gemini Managed Agents at AI Engineer – Jack Wotherspoon of Google DeepMind detailed Gemini managed agents. – These agents operate within cloud sandboxes using a single API call on Gemini infrastructure. – This method removes the need for complex containers, VMs, and networking setup. – The interactions API is designed to be agent-first, handling multi-step work like writing files. – The environment can be set to select a cloud sandbox on Google Cloud infrastructure. AI Engineer * Anthropic's CJ Avilla Details Vaults, Tool Allowlists for Running Many Agents – CJ Avilla of Anthropic outlined controls for running large numbers of AI agents. – Credentials are stored in vaults and injected only at network egress towards allowed hosts. – This prevents credentials from being exposed in uploaded environment files. – Default settings allow disabling tools like WebFetch and WebSearch. – The environment configuration utilizes a firewall to block traffic to unauthorized hosts. AI Engineer · at 1:27:46 · author has a stake: Anthropic * Heavy AI token spenders see 67% throughput gain, 32% lower cost per PR – Steve Jamison of Jellyfish analyzed company token spend per developer across quintiles. – The lightest fifth of companies spends only 0.3% of their PR cost on AI. – The heaviest fifth of companies spends 8.5% of their PR cost on AI. – The heaviest spenders achieved a 67 percent throughput gain and 32 percent lower cost per PR. – The bottom four-fifths of companies saw about 40 percent throughput gains and 25 percent lower cost per PR. AI Engineer · at 4:36 * Anthropic Staff Present Claude Managed Agents and Flows – CJ Avilla and Harrison Stall of Anthropic presented on Claude Managed Agents. – They discussed Anthropic's feature called Flows, which represents the next iteration of multi-agent orchestration. – Flow builds upon dynamic workflows currently available in Claude Code. – The presentation included a hands-on workshop featuring developer credits. AI Engineer · at 5:39 * Anthropic Adds Dynamic Workflows to Claude Managed Agents – Harrison Stall described Claude Managed Agents at the AI Engineer event. – The API allows developers to define agents using skills, tools, and MCP servers. – Developers can instantiate agents in a session and monitor a durable event log. – Multi-agent v1 launched in May, and dynamic workflows were shipped last Friday. AI Engineer · at 12:05 THE MAD PODCAST (MATT TURCK) * Mistral's Pierre Stock discusses Le Chonk and mixture-of-experts on The MAD Podcast – Pierre Stock, Mistral's VP of Science, discussed Le Chonk on The MAD Podcast. – Le Chonk is described as a mixture-of-experts model with a total of one trillion parameters. – The model utilizes 50 billion active parameters per token, matching the speed of a dense 50B model. – The architecture provides twenty times the capacity of a standard 50B model. – Stock cautioned that the model requires more data and careful training to prevent dead experts. The MAD Podcast (Matt Turck) · at 24:16 * Mistral VP Pierre Stock explains Le Chonk's single-model design and native multimodality – Pierre Stock of Mistral detailed Mistral Large 4, named Le Chonk, as a 1-trillion-parameter open-weight model. – The design allows one model to handle multiple reasoning modes, simplifying operation. – Native training on mixed text and images from scratch allows the model to learn domains like satellite images. – Mistral ships weights to customer tenants to enable sovereign inference. The MAD Podcast (Matt Turck) · at 36:06 * Mistral's Pierre Stock Describes Training Le Chonk on 3,800 GPUs – Pierre Stock stated that training the 1-trillion-parameter Le Chonk took months on 3,800 GPUs. – He used reinforcement learning, simulating a student taking exams 64 times. – The unstable asynchronous training was managed zero-shot. – Reward calibration involved defining rewards for correct answers and penalties for poor behavior. The MAD Podcast (Matt Turck) · at 45:48 * Mistral's Pierre Stock Details RL Training Domains for Le Chonk – Pierre Stock described five reinforcement learning domains for Le Chonk. – The coding domain uses fail-to-pass and pass-to-fail tests. – The finance and law domains rely on judge models and detailed rubrics. – The cyber domain involves painting attack scenarios for log analysis and patching. – Stock stated any verifiable sandbox task can be tackled using RL. The MAD Podcast (Matt Turck) · at 54:14 BIG TECHNOLOGY PODCAST (ALEX KANTROWITZ) * Anthropic's S-1 Filing Draws Emergency Pod Analysis – Dan Primack analyzed Anthropic's S-1 filing with Dan Primack of Big Technology. – Rittenhouse Research reported positive free cash flow of $3.4 billion in the first half of 2026. – Anthropic's revenue grew substantially from $41 million three years ago to $16.3 billion. – The filing reports $6.6 million in revenue per employee and 48 percent of revenue through hyperscalers. – Primack warned that committed compute contracts running to 2028 are difficult to reconcile with trailing figures. Big Technology Podcast (Alex Kantrowitz) · at 14:18 * Primack: Anthropic Cannot Afford to Withdraw Its S-1 – Dan Primack stated Anthropic needs capital and an IPO withdrawal would harm its reputation. – He cited the compute risk factor as the most practical concern in the filing. – Primack noted that Anthropic's trading multiples could set the benchmark for OpenAI's future IPO. – The S-1 reveals limited information regarding venture capital ownership beyond Spark's board seat. Big Technology Podcast (Alex Kantrowitz) · at 21:58 THE A16Z SHOW * Jassy: Amazon to Spend About $220 Billion on AI in 2026 – Amazon CEO Andy Jassy announced AWS grew 36.7 percent year over year last quarter on a $169 billion annual run rate. – The generative AI business at AWS exceeds $25 billion, up from $60 million three years ago. – Jassy projects that AWS will spend about $220 billion on AI in 2026. – This massive investment is driven by data center lead times and 30-year equipment useful lives. The a16z Show · at 4:11 THIS WEEK IN STARTUPS * OpenAI Engineer Arrested Over Alleged Boardroom Recording Device – The FBI arrested OpenAI engineer Abdul Rahman Al-Fazan. – Al-Fazan allegedly hid a recording device in an executive conference room. – Prosecutors found build instructions for the device in his wife's ChatGPT account. – The Saudi national booked a one-way flight to Saudi Arabia for October 12. This Week in Startups · at 1:21:16 CHINATALK * Ant Ling product lead calls AGI inevitable, urges better model evaluation – Richard Bian, product lead at Ant Ling, stated that AGI is inevitable. – He noted that current benchmarks like TOEFL and SAT are insufficient for effective model evaluation. – Bian argued that model testing is becoming a strategic process as the field expands into open ocean. – He described his year in the model lab as exceptionally long due to the rapid pace of innovation. ChinaTalk · at 15:08 OPENAI * OpenAI demos shared team tasks in ChatGPT – OpenAI released a demonstration of team tasks within ChatGPT. – This feature enables recurring jobs to run in the cloud for collaborative review. – Teammates can jointly edit instructions and review outputs from the shared task. – The walkthrough showed a legal team task automatically flagging a missing NDA company name. OpenAI --- The Observatory · https://observatory.mgorsky.net/archive/257/