AI News in October 2024: Swarm, Blackwell, Video Models and Edge AI

From New York, where IBM invited me to see its AI technology, I review a batch of announcements from October 2024. The video is sponsored by Mammouth. Product availability, prices and forecasts below describe that period rather than current offerings.

OpenAI has released Swarm, an open-source experimental framework for exploring multi-agent orchestration. Its repository describes it as educational, unsupported and not intended for production. I see it as another example of a platform provider moving into tools that developers might otherwise build themselves.

Swarm explores lightweight, customizable handoffs between agents. Unlike a hosted assistant service with built-in memory and retrieval, it runs mostly on the client and does not maintain state between calls. A weather example routes a request from a triage agent to a weather agent, which calls a tool and returns the result. Whether other models work depends on the integration; I have not tested that claim.

NVIDIA delivered an early DGX B200 system to OpenAI. I discuss a reported price around $400,000 and a system containing eight Blackwell GPUs, 1,440 GB of GPU memory, two Xeon CPUs and 4 TB of system memory. The displayed 3× training comparison is for a specified GPT-MoE workload against DGX H100; the cited 15× inference figure likewise depends on conditions, not every task.

NVIDIA also released Llama-3.1-Nemotron-70B-Instruct, a fine-tuned model. I have seen favorable benchmark claims and plan to test it myself. Reported comparisons against GPT-4o and Claude 3.5 Sonnet concern particular benchmarks and configurations, not a universal ranking.

Google announced an agreement to purchase power from Kairos Power’s planned small modular reactors to support energy demand, including AI. I favor nuclear investment but acknowledge I am not an expert. My claim that the US shut down all its nuclear plants after past accidents is incorrect: US nuclear plants continued operating.

Adobe announced Firefly video generation supporting text-to-video and image-to-video. Adobe described the model as designed for commercial safety based on its training-data approach; that is a company claim, not a guarantee about every output or legal use. I also mention Photoshop’s Generative Fill, which edits a selected area based on a prompt.

Worldcoin rebranded as World and introduced updates to its identity ecosystem. Its Orb scans a person’s iris as part of a proof-of-human system intended to help distinguish unique people from automated accounts. I say I recently interviewed founding member Saturnin Pugnet and discuss the project’s ambitions and potential uses.

My description that iris information is simply stored on a blockchain is incomplete. World announced personal custody in 2024 for data generated at the Orb, while its verification system uses additional cryptographic mechanisms. How biometric information is processed and protected is a separate question from whether a person can present a proof of human identity.

Pyramid Flow SD3 is another video-generation release with publicly available material on Hugging Face. I welcome an alternative to announced systems that were not yet widely available. Its model, code and use terms should be checked separately rather than assuming every component has the same unrestricted license.

Sierra, co-founded by former Salesforce co-CEO Bret Taylor, was reportedly raising funding at a valuation above $4 billion. It sells customer-service agents, including voice applications. The valuation and financing details are reported negotiations from that time, not a verified final deal in this video.

A report projected OpenAI losses rising to $14 billion in 2026. I interpret the figure in light of investment in staff, models, partnerships and data, but it was a forecast rather than a realized result. My prediction that the company could keep raising money is an opinion, not a financial guarantee.

Anduril introduced Bolt, a family of portable autonomous aircraft with reconnaissance and munition variants. The company describes Bolt-M as a precision system in which an operator makes engagement decisions. I find the technology interesting; the clip is a product announcement, not independent evidence of its operational performance.

SpaceX demonstrated a Super Heavy booster catch using the launch tower’s arms. I discuss the potential for faster reuse, but my suggestion of refueling and relaunching within hours is an aspiration, not a turnaround demonstrated by that flight. Earlier expendable boosters also do not describe every previous SpaceX landing.

Microsoft researcher Sébastien Bubeck reportedly left for OpenAI, illustrating the movement of talent between close partners that also compete. My statement that Microsoft owned about half of OpenAI is not substantiated by the video and should not be treated as a verified ownership figure.

Zyphra announced Zamba2-7B, a hybrid state-space model, with a chart comparing MMLU and time to first token against selected models. The chart is a vendor comparison under specific conditions, not proof that it wins on every task. I remain cautious until I can test it myself.

I discuss indications that web search features from the SearchGPT experiment could be integrated into ChatGPT. That was an expectation based on the information available in October 2024, not evidence of availability for every user at the time.

Mistral introduced Ministral 3B and 8B for edge use. The announcement describes up to 128K context and an interleaved sliding-window attention pattern for the 8B model. I favor small specialized models for many tasks, although my estimate that 98% of uses do not need a frontier model is not measured here.

Mistral’s published benchmark tables and historical API prices suggest particular cost and performance tradeoffs; they do not establish superiority on all tasks. The models had distinct commercial and research terms, so access to weights and permitted uses must be checked for the relevant version.

After Tesla’s event, reports indicated that Optimus robots had human assistance in some interactions. A subsequent video showed navigation in a factory, which Tesla presented as autonomous. A demonstration of one activity does not establish that every event interaction or robot capability was fully autonomous.

Power-over-Skin research imagines small wearables without their own batteries, including possible earbuds. I incorrectly describe this as using energy generated by the body: the researchers’ design transfers radio-frequency power through the body from a worn transmitter powered by another source. The project is research, not a demonstrated battery-free consumer earbud.

Google announced that NotebookLM Audio Overviews could be customized before generation and that a business offering was rolling out. I have not yet used the feature and am unsure how much I personally need audio summaries, though I recognize others find them useful.

Finally, paid ChatGPT users were receiving early access to a Windows app. I value native shortcuts for quickly starting a conversation and want easier access to ChatGPT or Perplexity on my phone. These rollout details reflect the time of the video.