AI News in September 2023: Apple, Model Releases and Infrastructure

In this September 15, 2023 roundup, Matthew Berman discusses consumer AI products, Apple, autonomous driving, enterprise announcements and safety. His forecasts and reports describe that week’s information, not a current assessment of the products or their capabilities.

Coca-Cola introduced Y3000, a flavor developed with human insights and AI assistance. Berman cites reporting that the company gathered people’s associations with the future, used AI to explore flavor pairings and generated imagery for a packaging mood board. He sees the release as an attempt to reach younger consumers.

Apple’s 2023 iPhone event emphasized features such as computational photography and predictive text without, in Berman’s account, foregrounding the term AI. He predicts a future version of Siri powered by Apple’s reported internal AI work and running largely on device. That was his speculation, not a feature Apple announced at this event.

Berman imagines a voice assistant that could complete multi-step tasks and discusses spatial video capture for iPhone 15 Pro and Apple’s forthcoming Vision headset. He also shows an object-capture example for creating 3D assets. His assistant scenario is a hypothetical, not a demonstration of those capabilities.

A WIRED report describes optional prison work in Finland labeling Finnish-language real-estate text for a startup, Metroc. Berman cites a reported three-hour shift and approximately $1.67 an hour, and weighs digital-work experience against unease about low-paid prison labor. The article’s example is a specific program, not evidence about prison labor generally.

At a Mexican congressional hearing, journalist Jaime Maussan presented two specimens and claimed they were non-terrestrial. The reported origin was near Cusco in Peru, not Mexico. Their extraterrestrial origin was not established, and assertions about unknown DNA do not establish alien life; Berman himself says the claims were unproven and likely false.

Stability AI introduced Stable Audio for text-to-audio generation. Berman calls the entire product open source and free, but the company’s September 2023 announcement described a free basic web tier and paid Pro tier; it did not establish that the original product’s model weights were open. Stable Audio Open was announced separately later.

Berman welcomes planned Slack features for summarizing threads, channel highlights and search, hoping they will help find important information in noisy conversations. He also discusses Anthropic’s then-new $20-per-month Claude Pro subscription, which promised higher usage and priority access. These are announcements and prices from 2023, not current availability or terms.

He cites a report that Character.AI had grown rapidly among younger mobile users, while saying he had not personally found much use for role-playing characters. The reported audience and usage estimates are attributed to that source and are not independently verified in the video.

Turning to Tesla, Berman describes reporting about a shift toward end-to-end neural-network driving and argues Tesla’s fleet data could be an advantage. He says his own test of the driver-assistance system was impressive but sometimes failed and required substantial attention. His claim that Tesla was far ahead of competitors is an opinion, not an independent safety comparison; drivers must supervise the system.

Berman cites reporting that Meta was assembling chips and infrastructure for a new model intended to compete with GPT-4, with training reportedly planned for 2024. These were reported intentions at the time, not a demonstrated result or a guarantee about a future model’s performance or licensing.

Nasdaq received SEC approval for its Dynamic M-ELO order type. According to Nasdaq, the system evaluates more than 140 market factors and adjusts a holding period by symbol; its reported test improvements concern its own fill-rate and market-impact measures. The approval did not mean every trade would avoid price increases or that all market orders used AI.

Microsoft announced a Copilot Copyright Commitment covering certain copyright claims for eligible commercial customers, including GitHub Copilot. Berman summarizes this as Microsoft assuming legal responsibility, but the commitment has conditions: customers must use the applicable safeguards and not deliberately seek infringing material. It is not blanket protection against every legal claim.

He discusses essay-writing services that advertise to students, including services combining AI output with human intervention. Berman compares AI writing tools with calculators and speculates that education may place greater emphasis on using them after learning basics. That is his educational view, not a conclusion that writing skills no longer matter.

Berman reports voluntary AI-safety commitments by additional technology companies and Google’s announced $20 million responsible AI fund. He welcomes the efforts while questioning whether voluntary pledges and the amount of funding are sufficient. A voluntary commitment is not the same as a binding legal guarantee.

He discusses Google’s disclosure requirements for certain AI-altered election ads and Amazon’s requirement that Kindle publishers disclose AI-generated content. Berman favors disclosure to audiences. He also mentions copyright disputes; whether AI-assisted material qualifies for copyright protection depends on human authorship and jurisdiction, rather than a single rule covering every use of AI.

Berman turns to water used for cooling data centers and cites an estimate of roughly 500 milliliters for a series of 5 to 50 prompts. That figure comes from a particular research estimate and depends heavily on location, cooling, power generation, model and workload; it is not a fixed water cost for every user or prompt.

An MIT researcher is developing algorithms to design concrete structures that use less material while preserving strength. Berman presents it as a promising research application; the example does not establish that every resulting design is ready for construction or that it achieves a specific emissions reduction.

He also describes Optellum’s software for assessing lung nodules using CT data, and suggests AI could help clinicians identify cancer earlier. Performance and clinical benefit depend on the population, validation and clinical workflow; the description is not evidence that the system can replace doctors or that every patient needs fewer scans.

Berman shares an AI-made recreation of a Pulp Fiction scene as his video of the week. He then discusses reports about YandexGPT and Russian-language performance, noting that he has not tested it and cannot substantiate a broad claim that it outperforms GPT-4. He closes by observing that AI development is a global competition.