Tech Brew Ride Home - Wed. 02/07 – AI Watermarks And New Apple AI Models

Starting point is 00:00:00 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 Tech Memerite Home for Wednesday, February 7th, 2020. I'm Brian McCullough today. Open AI is adding watermarks to Dolly 3 images. New AI model from Apple, new open source model that is king of the LLM Hill, more rumors of a foldable iPhone,

Starting point is 00:00:49 a new streaming service that will be like Hulu but for sports, and the specific words and phrases that will get your college application essay flagged as being AI generated. Here's what you miss today in the world of tech. Open AI says Dolly 3 will embed watermarking. into images with C2PA metadata, but acknowledges the metadata can be easily removed, either accidentally or intentionally. Quoting the Verge. The Coalition for Content Providence and Authenticity, a group consisting of companies like Adobe and Microsoft, has been pushing the use of the content credentials watermark to identify the provenance of content and show if it was made by humans or with AI.

Starting point is 00:01:31 Adobe created a content credential symbol, which OpenAI is adding to Dolly 3 creations. Meta recently announced it will add tags to AI-generated content on its social media platforms. OpenAI says watermarks from C2PA will appear in images generated on the chat GPT website and the API for the Dolly 3 model. Mobile users will get the watermarks by February 12th. They'll include both an invisible metadata component and a visible CR symbol, which will appear in the top left corner of each image. People can check the provenance, which AI tool was used to make the content of any image generated by OpenAI's platforms through websites like Content Credentials Verify. So far, only still images, not videos or texts,

Starting point is 00:02:14 can carry the watermark. But OpenAI points out that C2PA's metadata can, quote, easily be removed either accidentally or intentionally, especially as most social media platforms often remove metadata from uploaded content, taking a screenshot amidst the metadata. and quote. As mentioned, this follows news yesterday that meta has proposed standards to identify AI content that its industry peers could use and plans to label AI images posted on Facebook, Instagram, and threads as such. Sources are telling the information that Apple is building at least two iPhone prototypes that fold widthwise. The devices are in early development and not set for production in 2024- 2025, so we'll have to wait a while if we ever do see them.

Starting point is 00:03:04 quote, Apple recently approached at least one manufacturer in Asia for components related to two fultable iPhones of different sizes, they said. The products could be canceled if they don't meet Apple's standards, they said. Apple has explored foldable products for more than a decade, but its leader's interest in them has fluctuated, according to interviews with multiple former Apple employees and a review of its patent filings. Apple CEO Tim Cook began asking the company's designers and engineers about the possibility of a foldable iPhone as early as 2018. Also that year, he reacted positively to a demonstration by Apple's designers and engineers of a 7-inch foldable display, that person said. Two problems may stand in the way of a foldable iPhone.

Starting point is 00:03:43 Apple's engineers have struggled for years to overcome the technical challenges of building such a device, and its designers haven't come up with enough compelling features that would make consumers want one, especially given its high retail costs compared to non-foldable phones, according to three people with direct knowledge of the effort. former Apple employees said Apple's first foldable product would likely be the iPad, a lower-profile device that would test consumer appetite for a foldable iPhone. Apple has been working with South Korea's LG and Samsung, both of which have been involved for several years in making foldable displays for the foldable iPad prototypes, they said. Some details of a foldable iPad were first reported by Chinese, Taiwanese, and South Korean media, end quote. Apparently, the original vision for a

Starting point is 00:04:24 foldable iPhone had screens on the outside when closed, but this proved to be a non-starter. The primary concern for the hardware team was the device's susceptibility to damage from falls. They also aimed to create a foldable iPhone with a sleeker profile targeting a thickness reduction by 50% compared to existing models. However, this goal proved to be challenging due to the size constraints of batteries and display technologies. Around 2020, according to the sources, Apple shifted its focus from iPhones to the development of a foldable iPad, envisioned to be close in size to the iPad Mini. This change in direction was partly due to the device's use case.

Starting point is 00:05:00 The foldable iPad, potentially with an 8-inch screen, wouldn't need to fit into pockets or meet the rigorous drop-test standards of iPhones, allowing for a thicker design. An ex-Apple employee noted that engineers are currently tackling specific problems related to the persistent crease that appears with continuous folding. Additionally, there's an ongoing effort to refine the hinge mechanism to enable the device to lay flat when unfolded, eliminating any bumps or dips that would interfere with the functionality of the Apple Pencil. Sports was already shaping up to be

Starting point is 00:05:36 a catalyst for upending the later stages of the streaming wars, and now this has thrown a huge curveball into everything. Fox, ESPN, and Warner Brothers Discovery have announced a new joint venture to launch a streaming sports service in the U.S. in the fall of 2024. Each entity would own a third of the venture, quoting the journal. The as-yet unnamed service will be offered directly to consumers who would be able to stream all of these companies' sports content, the company said in a statement following a report in the Wall Street Journal about the new venture. For Disney, the partnership with other networks adds to an array of strategic options the company has explored for ESPN. Disney is still looking for a potential strategic partner or investor

Starting point is 00:06:18 and will maintain a plan to offer a standalone ESPN streaming app for those who don't want the all-on-one bundle from the three companies, people close to the situation said. There are risks to the tie-up. Disney knows as well. as any, the perils of a joint venture in media, it is now in the middle of trying to end its joint ownership of Hulu by buying out its partner Comcast after years of difficulties. Also, the new service won't include content from Paramount Global CBS or Comcast's NBC Universal. City analysts expect the new service to encompass about 55% of U.S. sports rights, according to a note published Tuesday. A chief executive for the venture is expected to be named in the coming weeks.

Starting point is 00:06:57 People familiar with the matter said, while no price tag has been set, it is expected to be significantly lower than the typical cable bundle, which often can run north of $100 a month, end quote. Yes, but don't expect this to come in at, say, $15 a month either. Given the fees that these companies already get from cable companies, ESPN commands around $11 a month per household, and you're adding two others here, and the pie gets split three ways. It's unlikely that this would retail at less than $40 a month, maybe even $50 a month. And naturally, three or so years later, after everybody is hooked on that content, that becomes $70 a month. But most of the sports is over here. You can see the potential attraction.

Starting point is 00:07:45 Another AI model from Apple has been released. Apple has partnered with University of California researchers to release open source AI model, MGIE, which can edit images based on natural language instructions, quoting Venturebeat. MGIE, which stands for MLLM guided image editing leverages multimodal large language models MLLMs to interpret user commands and perform pixel-level manipulations. The model can handle various editing aspects such as Photoshop style modification, global photo optimization, and local editing. M-G-I-E is based on the idea of using MLLMs, which are powerful AI models that can process both text and images to enhance instruction-based image editing. MLLMs have shown remarkable capabilities in cross-modal

Starting point is 00:08:33 understanding and visual-aware response generation, but they have not been widely applied to image editing tasks. M-G-I-E integrates MLLMs into the image editing process in two ways. First, it uses MLLMs to derive expressive instructions from user input. These instructions are concise and clear and provide explicit guidance for the editing process. For example, given the input, make the sky more blue, M-G-I-E can produce the instruction, increase the saturation of the sky region by 20%. Second, it uses MLLMs to generate visual imagination, a latent representation of the desired edit. This representation captures the essence of the edit and can be used to guide the pixel-level manipulation. MGI uses a novel end-to-end training scheme that jointly

Starting point is 00:09:19 optimizes the instruction derivation, visual imagination, and image editing modules. MGIE can handle a wide range of editing scenarios from simple color adjustments to complex object manipulations. The model can also perform global and local edits depending on the user's preference, end quote. This is on GitHub right now, so check it out. But obviously, if Apple's goal is to put language models on device on your phone, imagine what this could do for editing your Apple photos. Also, we have a new King of the Hill if the Hill is LLM Performance, and the new King is open source.

Starting point is 00:10:04 SMOG-72B is a Quen-72B-based open-source LLM released by Abacus A.I, and it has topped the Hugging Face Open LLM leaderboard by outperforming GPD3.5 on several benchmarks. Coding Venture Beat. Smag 72B is technically a fine-tuned version of Gwen 72B, another powerful language model that was released just a few months ago by Gwen, a team of researchers at Alibaba Group. What's most noteworthy about today's release is that Smog 72B outperforms GPD 3.5 and Mistral Medium two of the most advanced open-source large language models developed by OpenAI and Mistral,

Starting point is 00:10:43 respectively, in several of the most popular benchmarks. Smog-72B also surpassed Gwen 72B, the model from which it was derived by a significant margin in many of these evaluations. According to the Hugging Face Open LLM Leaderboard, which measures the performance of open-source language models on a variety of natural language understanding and generation tasks. Smog 72B is now the first and only open source model to have an average score of more than 80 across all major LLM evaluations. While the model still falls short of the 90 to 100 point average indicative of human level performance, its birth signals that open source AI may soon rival big tech's capabilities, which have long been shrouded in secrecy. In short,

Starting point is 00:11:22 the release of Smog 72B could fundamentally reshape how AI progress unfolds, tapping the ingenuity of those beyond just a handful of wealthy companies. With today's release, Smauk 72B becomes the first open source model to achieve an average score of 80 on the hugging face, open-l-lm-LM leaderboard, which is considered a remarkable feat in the field of natural language processing and open-source AI. Smok-S72B excels especially in reasoning and math tasks, thanks to the techniques that Abacus AI applied to the fine-tuning process. These techniques, which will be detailed in an upcoming research paper, target the weaknesses of large language models and enhance their capabilities, end quote. This year marks the first full college admissions season since the rise of Chachybt and the arrival of the AI moment generally.

Starting point is 00:12:11 College admissions officers are writing for a deluge of college admission essays written by AI robots. But the funny thing is, they claim they're not having that difficult at time spotting the AI-assisted writing. It can come down to the overuse or perhaps not quite apt use of specific words. What are the words that will flag your essay as possibly AI written? beacon, comprehensive curriculum, esteemed faculty, vibrant academic community, and, well, tapestry. This comes from Forbes, quote, Tapestry in particular is a major red flag in this year's pool, several essay consultants on the platform Fiverr told Forbes, Mike, an Ivy League alum and former editor-in-chief of the Cornell Business Journal, who now edits hundreds of grad school applications

Starting point is 00:12:58 each cycle through capital editors, said it's appeared repeatedly in drafts from at least 20 of his clients in recent months. He requested anonymity to protect their privacy. I no longer believe there's a way to innocently use the word tapestry in an essay. If the word tapestry appears, it was generated by chat GPT, he told Forbes. Though many such words on their own could have come from a human, when a trained eye sees them used over and over again in the same cadence across multiple essays, quote, it's just a real telltale sign. There will be a reckoning, Mike added. There are going to be a ton of students who unwittingly use the word tapestry or other words in their essay that may not be admitted this cycle, end quote. Ben Toll, the dean of undergraduate admissions at George Washington

Starting point is 00:13:42 University, explained just how easy it is for admissions officers to sniff out AI written applications. When you've read thousands of essays over the years, AI-influenced essays stick out, told Forbes. They may not raise flags to the casual reader, but from the standpoint of an admissions application review, they are often ineffective and a misused opportunity by the student, end quote. The words were thin, hollow, and flat, he said. While the essay filled the page and responded to the prompt, it didn't give the admissions team any information to help move the application towards an admit decision, end quote. This actually fits with my experience, generating text via chat GPT. I've often told people almost the opposite. No matter what you're trying to get it to

Starting point is 00:14:21 write, it ends up sounding like a sixth graders book report. You know, doll introduction, a few facts, dull, repetitive summary conclusion. In other words, thin, hollow, and flat. So I'm surprised to see AI's not great at writing in what seems to be its natural style. More phrases that the Forbes piece flags as dead giveaways, leadership prowess, commitment to continuous improvement and innovation, and structuring an argument by going, not merely X, but Y. The piece goes on to describe how a whole cottage industry of college application essay editors is now rising to the fore to make essays sound more human. Quoting again, Pajic, a freelancer in Serbia, who has three master's degrees and one Ph.D. has edited roughly 100 personal statements and college essays

Starting point is 00:15:07 through Fiverr. Though she initially feared her business would take a hit with the onset of chat GPT, the reality has been quite the opposite. The AI boom has fueled unprecedented demand for editors adept at making computer-generated writing sound like the real thing. ChatGPT is extremely wordy, she said, and there are many abstract words that don't really connect well. Today, she charges up to $100 to add human vibes to AI-written essays. Others offering similar services through Fiverr charge anywhere from $10 to $150, depending on the length, end quote. Nothing for you today. Talk to you tomorrow.

Tech Brew Ride Home - Wed. 02/07 – AI Watermarks And New Apple AI Models

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.