Tech Brew Ride Home - Wed. 09/11 – Mistral Goes Multimodel; And An AI Model Mystery

Episode Date: September 11, 2024

Mistral goes multimodal for the first time. Meta admits to scraping the data of every adult Australian. The details on the new PS5 Pro. Wouldn’t it be wild if, through stablecoins, crypto BECOMES the banking system instead of replacing it. And a weird mystery in AI land. Sponsors: Promevo.com/techmeme Links: Mistral releases Pixtral 12B, its first multimodal model (TechCrunch) Facebook admits to scraping every Australian adult user's public photos and posts to train AI, with no opt-out option (ABC News) The $700 PS5 Pro doesn’t come with a disc drive (The Verge) Exclusive Hands-On: I Played Sony's All-New PS5 Pro (CNET) Americans used record 100 trillion megabytes of wireless data in 2023 (Reuters) Payments in Singapore With Stablecoins Rise to Almost $1 Billion (Bloomberg) Reflection 70B model maker breaks silence amid fraud accusations (VentureBeat) 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 Tech meme right home for Wednesday, September 11th, 2020. I'm Brian McCullough.

Today, Mistral goes multimodal for the first time. Meta admits to scraping the data of every single adult Australian. The details on the new PS5 Pro, wouldn't it be wild if, through stablecoins, crypto becomes the banking system instead of replacing it, and a weird mystery in AI land. Here's what you miss today in the world of tech. French AI startup Mistrel has released its first multimodal model. Pextral 12B, available on GitHub and Hugging Face and via an API soon. Coding TechCrunch, called Pextral 12B.

The 12 billion parameter model is roughly 24 gigabytes in size. Parameters roughly correspond to a model's problem-solving skills, and models with more parameters generally perform better than those with fewer parameters. Built on one of Mistral's text models, Nemo 12B, the new model, can answer questions about an arbitrary number of images of an arbitrary size, given either URL's, or or images encoded using base 64, the binary-to-text encoding scheme. Similar to other multimodal models such as Anthropics' Claude family and OpenAISGPT40, Pextral 12B should, at least in theory, be able to perform tasks like captioning images

and counting the number of objects in a photo, available via a torrent link on GitHub and AI and machine learning development platform hugging face. Pextral 12b can be downloaded, fine-tuned, and used presumably under Mistral's standard dev license, which requires a paid license for any commercial applications, but not for research and academic uses. Mistral hasn't clarified exactly which license applies to Pixdrell 12B, however. The startup offers some models under an Apache 2.0 license without restrictions. We've reached out to Mistral for more information, and we'll update this post if we hear back. This writer wasn't able to take Pictrel 12B for a spin, unfortunately. There weren't any working web

demos at the time of publication. In a post on X, Sophia Yang, head of Mistrel, Developer Relations said Pixel 12B will be available for testing on Mistral's chatbot and API serving platforms lay chat and lay platform soon, end quote. Meta has admitted to scraping every Australian adult Facebook user's public data in order to use it to train AI models with also no opt-out option. Why were they able to do this? Because there ain't no law against it, sort of. Quoting ABC News, Meta's global privacy director, Melinda, Claybaw,

was pressed at an inquiry as to whether the social media giant was hoovering up the data of all Australians in order to build its generative artificial intelligence tools and initially rejected that claim. Labor Senator Tony Sheldon asked whether META had used Australian posts from as far back as 2007 to feed its AI products, to which Ms. Claibor responded, we have not done that. But that was quickly challenged by Greens Senator David Shoebridge. Shoebridge, quote, the truth of the matter is that unless you have consciously set those posts to private since 2007. Meta has just decided that you will scrape all of the photos and all of the texts from every public post on Instagram or Facebook since 2007 unless there was a conscious

decision to set them on private. That's the reality, isn't it? Claybaugh, quote, correct, end quote. Ms. Claibaw added that accounts of people under 18 were not scraped, but when asked by Senator Sheldon whether public photos of his own children on his account would be scraped, Ms. Claibaw acknowledged they would. In June, meta-notified users in the European Union and the United States that it would use their data to train its generative AI products such as meta AI unless users opted out. The company provided an opt-out option to EU users in part because of legal uncertainty surrounding strict privacy laws covering those nations. Ms. Claiborne admitted to the inquiry that those opt-out options were not offered to Australians. Quote, in Europe, there is an ongoing legal

question around what is the interpretation of existing privacy laws with respect to AI training. Ms. Claibaw said, we have paused launching our AI products in Europe while there is a lack of certainty. So you are correct that we are offering an opt-out to users in Europe. I will say that the ongoing conversation in Europe is the direct result of the existing regulatory landscape. Ms. Claibos said Australian users had the ability to set data to private, but opt-out options offered to Europeans were in response to laws enforced there. She said that Meta needed a lot of data in order to provide the most, quote, flexible and powerful AI tool it could, and that a lot of data was needed in order to deliver a safer product with fewer biases. The development comes a day after the federal

government vowed to introduce a ban on social media for children over concerns of harm the platforms were causing, end quote. Sony has announced the PS5 Pro offering a better GPU and faster memory for up to 45% faster game rendering, available from November 7th for $700. But warning, there will be no disc drive in this bad boy. Quoting the Verge, if you want to play your physical games on the new console, you'll have to pay extra. A disc drive isn't included, so you'll have to buy and install one separately for an additional $79.99. PS5 Pro is available as a disc-less console with the option to purchase the currently available disc drive for PS5 separately.

Sony's Hideki Nishino writes in a PlayStation blog post. That means if you are buying both the console and the disc drive together, you'll be paying a whopping $779.98. While very annoying, it sure seems like an expensive pro console should come with a disc drive. This move from Sony isn't entirely out of left field. Sony first released the separate PS5 disk drive alongside the slim PS5 redesign, giving you the option of adding a drive to an all-digital PS5 slim if you wanted that after the fact. Sony also sells a PS5 slim with the disc drive included, end quote. CNET already got hands-on with the thing, and here were their first impressions, quote,

I was shocked that the PS5 pro wasn't a hulking beast. The console's contours are nearly the same as the original PS5, and it's actually smaller. Meanwhile, the slim PS5 released last year is definitely smaller than the pro, but the difference in size isn't massive. As I jump back and forth, I can see the difference in gameplay. Everything is crisper, more fluid, or both. I'd prefer to play on the PS5 Pro, with a non-pro PS5 available for $500 today and likely less during upcoming holiday sales. I don't know if the sometimes subtle upgrades will be worth the price for many. The PS5 Pro is not the PlayStation 6, which likely won't be released for another 3.3.

or four years, and it isn't for everyone. It's a big, graphically boosted piece of hardware that can keep up with ever-changing PCs, and in some ways, maybe exceed them. The PS5 Pro is all about making big TV gaming a happier experience. A souped-up GPU promises more ray tracing and fluid 4K and 60 frames per second gaming across the board for games that get pro-upgraded and automatic AI upscaling for other games in the PlayStation Library. Oh, and expect a lot more ray tracing, a fancy graphical technique to simulate light. Expect games to get pro-upgraded performance extras. Sony says about 40 to 50 games will get patches when the system launches in November. Everything the PS5 Pro offers is about graphics. The CPU is the same as the PS5, and so is the SSD speed. The GPU, meanwhile, has 67% more

computing cores, according to Sony, with 28% faster RAM and 45% faster rendering. There are three big initial upgrade, Sony is specifically pushing on the pro for what that new GPU is doing, more ray tracing, automatic AI-assisted game upscaling for 4K, and a new pro mode for games that will combine 60 frames per second and 4K together, end quote. Just an interesting number here. Americans used 100 trillion megabytes of wireless data last year, quoting Reuters. That was up 36% over the prior year in the largest single-year increase in wireless data consumption, according to an industry survey released. on Tuesday. The increase, 26 trillion megabytes over 2022, comes as a growing number of 5G wireless

devices are being used, said Wireless Industry Association CTIA that represents major wireless carriers like Verizon, AT&T, T-Mobile, and technology firms. The total number of wireless connections rose to 558 million last year, up 6% over 2022, the survey found. Demand for spectrum use is soaring, driven in part by more wireless use and advancements, including drones, self-driving vehicles, space missions, and precision agriculture. The survey said the number of minutes Americans spent talking on the phone fell slightly from $2.5 trillion in 2022 to $2.4 trillion in 2023, and text messages were about the same at $2.1 trillion in 2023 over the prior year. The rise in wireless comes amid a standoff in Congress over how to find new spectrum. Congress in March 2023 let the Federal

Communications Commission authority to auction spectrum lapse for the first time in three decades, amid debate about what spectrum used by the Defense Department could be repurposed or shared. There is no pipeline of spectrum for 5G, said CTIA CEO Meredith Atwell Baker. To continue to meet the insatiable demand for wireless drive innovation and support America's economic competitiveness, the wireless industry needs access to more full power licensed spectrum, end quote. The Biden administration announced steps in November 2020, aimed at freeing up additional wireless spectrum by repurposing spectrum currently set aside for parts of the federal government, but has come under fire from Republicans for not moving fast enough, end quote.

There was an odd lots episode last week, the podcast Oddlots, that made the case that we've sort of discussed before. All of the sudden, the big winners in the crypto space seem to be stablecoin companies. The guest on that Oddlots show made the case that weirdly, this form of crypto actually plugs holes in the existing international banking system that could use a fix. So could stablecoins somehow become some sort of underlying scaffolding of the international monetary system? That would be ironic. Well, here's an interesting data point. According to Chainalysis, payments in Singapore using Stablecoins reached a record high of almost $1 billion in Q2, led by transactions at merchant outlets.

Quoting Bloomberg, businesses use the tokens because of efficiency and low cost. Jainalysis' cybercrimes research lead Eric Jardine said, Staplecoin payments in Singapore amounted to about $161 million in the second half of 2023, the study showed, so that's quite a leap in six months. Stablecoins typically seek to hold a steady value of $1 backed by reserves of cash and bonds. They are mostly used for crypto trading, and some have been faulted for their popularity with the criminal underworld. Stable coins remain a small sliver of payment flows, for example, retail card payments

in Singapore were worth $73.2 billion. in Singapore money or $56.2 billion in the second half of last year. The city-state is trying to develop a digital asset hub focused on institutional uses of blockchain technology that make financial payments quicker and cheaper, end quote. Finally today, a bit of a mystery back in AI land. Matt Schumer has been accused of fraud over hyper-rights 70B parameter AI model. He now says he, quote, got ahead of himself, but doesn't explain why the model he was hyping underperform. formed when others tested it out. Quoting Venture Beat. Matt Schumer, co-founder and CEO of OtherSide AI, also known as its signature AI assistant writing product, HyperWrite, has broken his

near two days of silence after being accused of fraud when third-party researchers were unable to replicate the supposed top performance of a new large language model he released on Thursday, September 5th. On his account on the social network X, Schumer apologized and claimed he, quote, got ahead of himself, adding, I know that many of you are at a lot. excited about the potential for this and are now skeptical, end quote. However, his latest statements do not fully explain why his model, Reflection 70B, which he claimed to be a variant of Mehta's Lama 3.1, trained using synthetic data generation platform, glave AI, has not performed as well as he originally stated in all subsequent independent tests, nor has Schumer clarified precisely

what went wrong. In case you're just catching up, last week, Schumer released Reflection 70B on the open source AI community hugging face, calling it, quote, the world's top open source model. in a post on X and posting a chart of what he said were at state-of-the-art results on third-party benchmarks. Schumer claimed the impressive performance was achieved thanks to a technique called reflection tuning, which allows the model to assess and refine its responses for correctness before outputting them to users. However, just days after its debut and over the last weekend, independent third-party evaluators and members of the open-source AI community posting on Reddit and Hacker News began questioning the model's performance and were unable to replicate it on their own.

Some even found responses and data indicating the model was related to perhaps merely a thin wrapper, pointing back to Anthropics Claude 3.5 Sonnet model. Criticism mounted after artificial analysis and independent AI evaluation organization posted on X that its test of Reflection 70B yielded significantly lower scores than initially claimed by Hyperite. Also, Schumer was found to be invested in Glave, the AI startup he said, whose synthetic data he used to train the model on, which he did not disclose when releasing reflection 70B. Schumer attributed the discrepancies to issues during the model's uploading process to hugging face and promised to correct the model waits last week, but has yet to do so.

One ex-user, Shin Megami-B, Bawson, openly accused Schumer of, quote, fraud in the AI research community on Sunday, September 8th. Schumer did not directly respond to this accusation. After posting and reposting various X messages related to Reflection 70B, Schumer went silent on Sunday and did not respond to Ventrabeats request for comments, nor post any public ex-posts until this evening, Tuesday, September 10th. Schumer finally released a statement on X at 5.30 p.m. Eastern Time, apologizing and stating in part, we have a team working tirelessly to understand what happened and will determine how to proceed once we get to the bottom of it. Once we have all the facts, we will continue to be transparent with the community about what happened and next steps,

end quote. Schumer also linked to another ex post by Sahil Chaudry, founder of Glave AI, the platform Schumer previously claimed was used to generate synthetic data to train Reflection 70B. intriguingly, Chowdry's post stated that some of the responses from Reflection 70B saying it was a variant of Anthropics Claude are also still a mystery to him. He also admitted that, quote, the benchmark scores I shared with Matt haven't been reproducible so far. However, Schumer and Chowdry's responses were not enough to mollify skeptics and critics, including Yunchin Jin, co-founder and chief technology officer of hyperbolic labs and open access AI cloud provider. Jin wrote in a lengthy post on X detailing how hard he worked to post a version of Reflection 70B on his site, and troubleshoot the supposed errors, noting that, quote,

I was emotionally damaged by this because we spent so much time and energy on it, so I tweeted about what my faces looked like during the weekend, end quote. He also responded to Schumer's statement with a reply on X, writing, Hi, Matt, we spent a lot of time, energy, and GPUs on hosting your model, and it's sad to see you stopped replying to me in the past 30-plus hours. I think you can be more transparent about what happened, especially why your private API has a much better Perf, end quote. McGammy Boston, among many others, remained unconvinced as of tonight,

in Schumers and Chaudry's telling of events and casting the saga as one of mysterious, still unexplained errors borne out of enthusiasm. As far as I can tell, either you are lying or Matt Schumer is lying, or, of course, both of you. He posted on X, following up with a series of questions, end quote. Nothing more for you today. Talk to you tomorrow.