Tech Brew Ride Home - Thu. 07/20 – An Apple GPT

Episode Date: July 20, 2023

Rumors that Apple already has an LLM framework and their own generative chatbot. Google is shopping an AI bot that can write news stories to various journalism outlets. Has GPT-4 actually gotten dumber of late? More streaming price raises. And it’s the last chance saloon for Microsoft’s AR headset for the US Military. Sponsors: BirdDogs.com/ride notion.com/ride Links: Apple Tests ‘Apple GPT,’ Develops Generative AI Tools to Catch OpenAI (Bloomberg) Google Tests A.I. Tool That Is Able to Write News Articles (NYTimes) Study claims ChatGPT is losing capability, but some experts aren’t convinced (ArsTechnica) Meta open-sources Llama 2, but with strings attached (MoneyControl) Apple slams UK surveillance-bill proposals (BBC News) Google raising price of YouTube Premium to $13.99 per month (9to5Google) Microsoft Poised to Deliver Improved Combat Goggles, US Army Says (Bloomberg) Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the TechMeme right home for Thursday, July 20th, 20th, 203, I'm Brian McCullough today. Rumors that Apple already has an LLM framework and their own generative chatbot.

Google is shopping an AI bot that can write news stories to various journalism outlets. Has GPT4 actually gotten dumber of late? More streaming price raises. And it's the last chance saloon for Microsoft's AR headset for the U.S. military. Here's what you miss today in the world of tech. I guess you could file this under the, of course they are file, with maybe a subheading of the why wouldn't they file, but it's news nonetheless that Mark German's sources are telling him that Apple has already built Ajax, a framework for building LLMs, and they have even created

a chatbot app that some engineers are calling Apple GPT. Rumor is that Apple is planning some sort of AI announcement next year. And again, why not? Because another. A.I. Announce another company gets an all-time high stock price. Quoting Mark, in recent months, the AI push has become a major effort for Apple with several teams collaborating on the project, said the people who asked not to be identified because the matter is private. The work includes trying to address potential privacy concerns related to the technology. Apple shares gained as much as 2.3% to a record high of $198.23 a share after Bloomberg reported on the AI effort Wednesday. Behind the scenes, Apple has

grown concerned about missing a potentially paramount shift in how devices operate. Generative AI promises to transform how people interact with phones, computers, and other technology. And Apple devices, which produced revenue of nearly $320 billion in the last fiscal year, could suffer if the company doesn't keep up with AI advances. That's why Apple began laying the foundation for AI services with the Ajax framework, as well as a chat GPT-like tool for use internally. AJAX was first created last year to unify machine learning development at Apple, according to the people familiar with the effort. The company has already deployed AI-related improvements to search, Siri, and maps based on that system. And Ajax is now being used to create large language models and serve

as the foundation for the internal chat GPT-style tool that people said. The chatbot app was created as an experiment at the end of last year by a tiny engineering team. Its rollout within Apple was initially halted over security concerns about generative AI, but has since been extended to more employees. Still, the system requires special approval for access. There's also a significant caveat. Any output from it can't be used to develop features bound for customers. Even so, Apple employees are using it to assist with product prototyping. It also summarizes text and answers questions based on data it has been trained with. Apple employees say the company's tool essentially replicates Bard, chat GPT, and Bing AI, and doesn't include any novel features or technology. The system is accessible as a web

application and has a stripped-down design not meant for public consumption. As such, Apple has no current plans to release it to consumers, though it is actively working to improve its underlying models. Beyond the state of the technology, Apple is still trying to determine the consumer angle for generative AI. It's now working on several related initiatives, a cross-company effort between its AI and software engineering groups, as well as the cloud services engineering team that would supply the infrastructure for any significant new features. While the company doesn't yet have a concrete plan, people familiar with the work believe Apple is aiming to make a significant AI-related announcement next year. Around the same time that it began developing its own tools, Apple conducted

a corporate trial of OpenAI's technology. It also weighed signing a larger contract with OpenAI, which licenses its services to Microsoft, Shutterstock, and Salesforce. Apple's AJAx system is built on top of Google Jacks, the Search Giant's Machine Learning Framework. Apple's system runs on Google Cloud, which the company uses to power cloud services alongside its own infrastructure and Amazon's AWS. As part of its recent work, Apple is seeking to hire more experts in generative AI. On its website, it is advertising for engineers with a, quote, robust understanding of large language models and generative AI and promises to work on applying that technology to the way, quote, people communicate, create, connect, and consume media on iPhones and its other devices, end quote.

sources are telling the Times that Google is testing an AI tool to produce news story, pitching this service to The Times, as well as the Washington Post News Corp and other outlets as a helpmate for journalists. Quote, the tool known internally by the working title Genesis can take in information, details of current events, for example, and generate news content the people said, speaking on the condition of anonymity to discuss the product. One of the three people familiar with the product said that Google believed it could serve as a kind of personal assistant for journalists, automating some tasks to free up time for others, and that the company saw it as responsible technology that could help steer the publishing industry away from the pitfalls of generative AI.

Some executives who saw Google's pitch described it as unsettling, asking not to be identified discussing a confidential matter. Two people said it seemed to take for granted the effort that went into producing accurate and artful news stories. news organizations around the world are grappling with whether to use artificial intelligence tools in their newsrooms. Many, including the Times, NPR, and Insider have notified employees that they intend to explore potential uses of AI to see how it might be responsibly applied to the high-stakes realm of news, where seconds count and accuracy is paramount. But Google's new tool is sure to spur anxiety, too, among journalists who have been writing their own articles for decades. Some news organizations, including the Associated Press, have long used AI to generate stories about matters including corporate earnings reports, but they remain a small fraction of the services articles

compared with those generated by journalists. Artificial intelligence could change that, enabling users to generate articles on a wider scale, that, if not edited, and checked carefully, could spread misinformation and affect how traditionally written stories are perceived, end quote. But wait, a new study is claiming that GPT-4's outputs became worse between March and June of this year. but we can't actually be sure because OpenAI is so opaque with its AI models. Quoting Ars Technica, in a study titled, How is Chat GPT's behavior changing over time, published on Archive, Ling Zhao Chen, Matezah Zaharia, and James Zau, cast doubt on the consistent performance of OpenAI's large language models,

specifically GPT3.5 and GPT4, using API access. They tested the March and June 2023 versions of these models on tasks like math problem solving, answering sensitive questions, code generation, and visual reasoning. Most notably, GPT4's ability to identify prime numbers reportedly plunged dramatically from an accuracy of 97.6% in March to just 2.4% in June. And even more strangely, GPT 3.5 showed improved performance in the same period. This study comes on the heels of people frequently complaining that GPT4 has subjectively declined in performance over the past few months. Popular theories about why include open AI distilling models to reduce their computational overhead and a quest to speed up the output and save GPU resources,

fine-tuning additional training to reduce harmful outputs that may have unintended effects, and a smattering of unsupported conspiracy theories such as OpenAI reducing GPT4's coding capabilities, so more people will pay for GitHub copilot. Meanwhile, OpenAI has consistently denied any claims that GPT4 has decreased in capability. As recently as last Thursday, OpenAI VP of product Peter Weelander tweeted, quote, No, we haven't made GPT4 dumber quite the opposite. We make each new version smarter than the previous one. Current hypothesis, when you use it more heavily, you start noticing issues you didn't see before, end quote.

While this new study may appear like a smart, smoking gun to prove the hunches of the GPT4 critics, others say not so fast. Princeton computer science professional Arvin Narayanan thinks that its findings don't conclusively prove a decline in GPD4's performance and are potentially consistent with fine-tuning adjustments made by OpenAI. For example, in terms of measuring code generation capabilities, he criticized the study for evaluating the immediacy of the code's ability to be executed rather than its correctness. The change they report is that the newer GDRICK, GPD4 adds non-code text to its output. They don't evaluate the correctness of the code. Strange,

he tweeted. They merely check if the code is directly executable, so the newer models attempts to be more helpful counted against it, end quote. While the paper by Chen, Zaharia and Zhao may not be perfect, Willison sympathizes with the difficulty of measuring large language models accurately and objectively. Time and again, critics point to open-eye's currently closed approach to AI, which for GPD4 did not reveal the source of training materials, source code, neural network weights, or even a paper describing its architecture. With a closed black box model like GPT4, researchers are left stumbling in the dark trying to define the properties of a system that may have additional unknown components, such as safety filters or the recently rumored eight

mixture of experts' models working in concert under GPT4's hood. Additionally, the model may change at any time without warning. One solution to this developer instability and researcher uncertainty may be open source or source available models such as Meta's Lama. With widely distributed weights files, the core of the model's neural network data, these models can allow researchers to work from the same baseline and provide repeatable results over time without a company like OpenAI, suddenly swapping models or revoking access through an API, end quote. Speaking of open source, folks have noticed that in Lama 2's commercial terms,

META says companies with more than 700 million monthly active users must request a license, and users are prohibited from utilizing LMA II to improve other LLMs. So number one, is this an attempt to keep folks like, I don't know, Snap from using it to improve their services, and number two, is this an attempt to keep folks like OpenAI from using Lama to improve GPT number whatever, quoting money control? Although Meta has announced that it will open source its LLM and allow free usage for commercial and research purposes, a closer examination of its community license agreement reveals a different story. According to Meta's commercial terms, companies with over 700 million or more monthly active users are required to obtain a license for Meta.

This means that meta's new AI technology is off limits to some of its social media competitors. Additionally, another clause in the commercial terms of Lama 2 states that users are prohibited from utilizing it to enhance or improve other large language models, aside from Lama 2 itself. Meta's decision to open access its AI models aligns with a broader trend in the industry. Abu Dhabi-based Technology Innovation Institute recently released Falcon LLM, making its code freely available this year. Additionally, Mosaic ML, which was acquired by Databricks for approximately $1.3 billion last month, also provides open source software for training LLMs, end quote. Non-AI news, if you can believe it. Apple is threatening to remove FaceTime and I message from the UK rather than weaken their security to comply with a proposed security bill amendment dubbed a snoopers charter.

Quoting the BBC, the government is seeking to update the investigatory powers or IPA 2016 law. It wants messaging services to clear security features with the home office before releasing them to customers. The Act lets the home office demand security features are disabled without telling the public. Under the update, this would have to be immediate. it. Currently, there has to be a review. There can also be an independent oversight process and a technology company can appeal before taking any action. Because of the secrecy surrounding these demands, little is known about how many have been issued and whether they have been complied with. But many messaging services currently offer end-to-end encryption so messages can be unscrambled only by

the devices sending and receiving them. WhatsApp and Signal are among the platforms that have opposed a clause in the online safety bill allowing the communications regulator to require companies to install technology to scan for child abuse material in encrypted messaging apps and other services. They will not comply with it, they say, with signal threatening to walk from the UK. Apple has also opposed the plan. The government has opened an eight-week consultation on the proposed amendments to the IPA, which already enables the storage of internet browsing records for 12 months, and authorizes the bulk collection of personal data. They are not about the creation of new powers, but making the act more relevant to current technology. The government says,

Apple says it has consistently opposed the act, originally dubbed a Snoopers' Charter, by critics, end quote. Who says inflation is coming down? Certainly not seeing that in the streaming arena. More price raises there. YouTube has increased the price of a premium subscription by $2 to $13.99 per month in the U.S. and has raised the YouTube music subscription from $999 to $10.99 per month. Quoting 9 to 5 Google. Toward the end of last year, family premium plan saw a big hike to $22.99 per month. That remains the same today. The annual subscription, which was introduced in January of 2022, goes to $139.99 in a $20 increase. Compared to paying monthly, you save $27.89.

YouTube last raised the price of YouTube premium, previously called Red, in 2018, with the relaunch of YouTube music, end quote. Finally, today, the U.S. Army says Microsoft plans to deliver improved goggles, based on the HoloLens to it by July 31st for intensive testing with real soldiers ahead of a decision on whether or not to go ahead with their deployment on potential battlefields or to full-on cancel the entire effort. Quoting Bloomberg, after delivery, the first 20 prototype IVAS 1.2 goggles will be assessed by two squads of soldiers in late August to check for improvements in reliability, low-light performance, and how well they fit soldiers without repeats of the nausea and dizziness that halted the deployment

of earlier versions. Microsoft said in a statement that the deliveries will be three months ahead of schedule. Over a decade, the Army projects spending as much as $21.9 billion for as many as 121,000 devices, spares, and support services if all options are exercised. If the initial assessment is a success, the Army will award a contract between July and September of next year to produce additional devices. for a second soldier evaluation. Success there would be followed by a full-blown combat operational test between April and June of 2025, determining whether the goggles could be deployed to combat

units within months, end quote. We had that crunch meeting last night about our AI project, and a few more tweaks are needed before we can start testing it, but actually, it became clear to me that we will need more testers than I thought as recently as yesterday. Maybe not all 300 of you, but certainly more than the dozen that I expected to have to make use of. As soon as I get this show out this afternoon, I'm going to run a few final tests myself, and then some of you might be getting some emails from me as soon as tonight. By the way, even if you don't test it for me, once I bring this experiment out of beta, I'll tell you all about it. Don't know. Feels like we might have something useful here. Talk to you tomorrow.