GPT-4.5 - Does a Cached Announcement Blog Prove It’s Coming?

Today on the AI Breakdown. The European Parliament has passed the AI act before that on the brief Does a Cached blog suggest that GPT 4.5 is coming? The AI breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our Discord, our YouTube and our newsletter. Welcome back to the AI Breakdown Brief. All all the AI headline news you need in around five minutes. Questions around OpenAI's next model have reached an absolute fever pitch recently. In general, people have been waiting for a very long time for OpenAI to release whatever is coming next, given that their GPT4 has been the standard bearer for more than a year. However, what has really increased focus on when OpenAI's next model is going to come is of course the release of Claude 3, which is the first time that there is broad consensus that a model is actually better than GPT GPT4 in practice. Sure, Google's Gemini Ultra claimed to be better uncertain benchmarks, but Claude 3 is really the one that I'm seeing people actually switch their behavior to. Take that plus Elon Musk's lawsuit and you have a fierce little cauldron of rumor and innuendo. One of those rumors exploded yesterday when people started reporting that they were finding a cached blog post showing up in search results on DuckDuckGo and on Bing. It was theoretically a blog post on OpenAI.com, where where the preview reads OpenAI announces GPT 4.5 Turbo, a new model that surpasses GPT 4 Turbo in speed, accuracy and scalability. Learn how GPT4.5 can generate natural language or code with a 256k context window and a knowledge cutoff of June 2024. You might notice that June 2024 is in the future from where we are now, leading to speculation that the plan was to be training it up until then and then releasing it. More than a lot of rumors. I saw numerous people tweeting that they had repeated this and they had also found it some used copilot to try to get more information out of that cached piece. And keep in mind this is from not necessarily normal hypey sources. Still, some people aren't buying it. Matt Popovich, the co founder and CEO of legislature AI says really doubting the GPT 4.5 turbo cached Bing result means anything. 1. There's hints 4.5 is being skipped over in favor of 5. 2. I doubt they'd release a turbo version of a brand new model substantially more powerful than GPT 4. 3. The cached page is over 5 months old now on that 4.5 being passed over for 5 point some of the folks best associated with OpenAI leaks, like the Flowers from the Future account have been on this tip that we will not see a GPT 4.5 but will instead go straight to GPT 5. On March 9, Flowers tweeted, no, GPT 5 is not coming next week and GPT 4.5 is still canceled. As I said. That didn't stop Flowers, however, from saying that this will be an incredible week that will shock the entire industry, but it won't be GPT 4.5 or GPT 5. Given how much chatter there has been about Cognition's Devon, the AI coding platform, maybe that's what they're referring to. In any case, I think that the sourcing is about as accurate as this tweet from Accelerate Harder, who writes, it was revealed to me in a dream that GPT 4.5 is coming. Basically, where we are right now is that there is a very broad and strong sense that OpenAI has to do something that for the first time since we've all been paying attention, they are plausibly behind and getting more so every day. And I think that a lot of the energy around the discussion shows the general anticipation and assumption that something big is coming. Meanwhile, moving over into Google land for a minute, the company has announced some new policies around the 2024 elections. Specifically, it is now restricting Gemini from answering Any questions about 2024 Elections in countries that are holding those elections this year, including the US And India. The company's India team writes, out of an abundance of caution on such an important topic, we have begun to roll out restrictions on the types of election related queries for which Gemin responses. This is an update from a policy announced in December that said it would be limiting election related queries, obviously ratcheting things up even a little bit more. Although this started in India, TechCrunch confirmed with Google that it was going to roll out these restrictions globally. Now, some argue that this should give us pause about Gemini's capacity to give us good information in the first place. Cornell professor Daniel Suesser writes, if Google's generative AI tools are too unreliable for conveying information about democratic elections, why should we trust them in other contexts such as health or financial information? What does that say about Google's long term plans to incorporate generative AI across its services, including search? From where I'm sitting, the more notable response is actually the lack of response this is the type of thing that I think a month ago would have provoked a big outcry on Twitter x about Google censoring things, et cetera, et cetera. But in the wake of the dust up around Gemini's image inaccuracy, it feels more and more like this is kind of assumed behavior and that people are, at least for now, writing Google off. That may just be a reflection of the moment, or it may be that people just aren't that interested in election policies, but it does feel to me at least a little bit notable. Even while big companies are restricting their tools, this week has seen a raft of startup announcements. One of those companies is called Physical Intelligence, and here's how one of the founders, Lockheed Groom, describes it. Lockheed there's been so much progress in AI, but it's been largely constrained to the digital world. At Physical Intelligence, we're developing foundation models to power any and all physically actuated devices with an early focus on dexterous manipulation. We're spending most of our time on scale data collection, algorithmic development and training large scale models. We are extremely early in setting out to create a world where atoms are almost as easy to manipulate as bits. I'm excited to share a lot more soon. Well, one of the things that was shared was that the company is well resourced to go after this mission. Physical intelligence, or PI, has raised $70 million in a pre seed round to do this. It is set to work creating its own AI model designed to bring basic human abilities to machines. But what fundamentally makes humans interesting is the brain, not our hardware. If you've been listening to this show, you know that the robotics AI overlap is heating up in a significant way and it's pretty interesting to see how this company is approaching those problems a little bit differently. For now, however, that is going to do it for today's AI Breakdown brief. Up next, the main AI breakdown. The tools that help us manage our projects and endeavors become way more significant feeling than most other normal tools we interact with. They are the things that help us breathe life into ideas. And that's why I am so pleased that today's episode is brought to you by Notion. Notion is the power center for all of my projects, from the AI Breakdown to the new education initiative that so many of you are participating in. Notion combines your notes, documents and projects into one space that's simple and beautifully designed. They've also now brought the power of AI directly into the platform so that when it comes to writing, brainstorming, getting creative with new ideas, strategic planning, all of these things where AI can be an incredible companion. In Copilot, all of that is available to you without leaving Notion. Now, Notion is a tool for you, but it's also for your teams. In fact, the more your team uses it, the more useful it becomes for everyone. Notion is used by over half of Fortune 500 companies and teams that use Notion send less email, cancel more meetings, save time searching for their work, and reduce spending on tools, which helps keep everyone on the same page. Try Notion for free when you go to notion.com AIbreakdown that's all lowercase letters notion.com AIbreakdown to try the powerful, easy to use Notion AI today. And when you use our link, you're supporting the show notion.com aibreakdown welcome back to the AI Breakdown. The breaking news this morning is that the European Parliament has passed the AI Act. I couldn't help but laugh when I saw a meme from Kier Lamont on Twitter who wrote how the EU AI act news sounds to us non EU legislative process understanders with an image that says, rest assured, this will be the sixth time we have passed the EU AI act and we have become exceedingly efficient at it. The joke here is, of course, that it is a very complex process to get legislation into law in the European Union. But this was in fact a significant and important step. Back in December, the European Union had gotten to provisional political consensus, which was itself not the first time that they had voted on this set of rules. But in Wednesday's session, the European Parliament actually endorsed the full Act. The final tally was 523 votes in favor, 46 against, and 49 votes abstaining. Now for a little bit of background. If you were just coming to this conversation, the AI act did not, as one might have guessed, start with the initiation of ChatGPT. In fact, they have been negotiating this law since 2021 and indeed one of the big challenges for them in terms of finalizing it was how to deal with generative AI, given that it hadn't been part of the early negotiations. The central underlying premise of the EUI act is to divide artificial intelligence into different categories based on risk. Louisa Juravsky on Twitter did a good breakdown, writing, the AI act follows a risk based approach. Some AI systems, she points out, are banned, including those that involve cognitive behavioral manipulation of people or specific vulnerable groups social scoring, that is Classifying people based on behavior, socioeconomic status or personal characteristics biometric identification and categorization, real time and remote biometric identification systems such as facial recognition. You may notice in other words, that what the European Union has classified as the riskiest potential applications of AI are basically the skynet totalitarian surveillance types of applications. The next category down is a high risk category which is fairly broad. Some of the systems that fall into that category include critical infrastructure like transport that could put the life and health of citizens at risk, education or vocational training that may determine access to education and professional course of someone's life, for example scoring of exams, safety components of products. The example given being AI application in robot assisted surgery, employment management of workers and access to self employment, think CV sorting software for recruitment, essential private and public services, credit scoring, denying citizens opportunity to obtain a loan, law enforcement that may interfere with people's fundamental rights, for example, the evaluation of the reliability of evidence, migration, asylum and border control management, for example, automated examination of visa applications and administration of justice and democratic processes like AI solutions to search for court rulings. So of course, the common thread here is that these are all applications where the stakes of what happens or what is decided is extremely high for the person whose life is affected by that decision. It's not just physical harm or risk, but things that could impact someone's career, someone's livelihood, someone's legal treatment. And so what does it mean to be in the high risk category? Well, it doesn't mean that things are banned like the highest risk category that we were just discussing. Instead, these high risk systems have to be assessed before being put on the market and throughout their life cycle. What's more, there will be specific procedures and processes for people to file complaints about how these systems are being used. What about generative AI? This was a big sticking point and the subject of a lot of intense lobbying and negotiation. These types of tools will not be classified as high risk, but will have their own set of rules as per the AI Act. Specifically, they'll have to comply with certain transparency requirements in EU copyright law, such as things like disclosing that content was generated by AI, an obligation to design models to prevent them from generating illegal content, and publishing summaries of the copyrighted data that was used for training. The Wall Street Journal describes some additional requirements around generative AI. Quote, makers of the most powerful AI models deemed to have what the EU calls a systematic risk will be required to put those models through state of the art safety evaluations and notify regulators of serious incidents that occur with their models. They will also have to implement mitigations for potential risks in cybersecurity protection. The provisions on general purpose AI were added during the EU's legislative process. Industry groups and some European governments pushed back against the introduction of blanket rules for general purpose AI, saying legislators should focus instead on risky uses of the rather than the models that underpin its use. France, home to Mistral AI and Germany sought to water down some of the legislation's proposals. Mistral's chief executive Arthur Mensch said recently that the AI act would, after some changes in final negotiations that lighten some obligations, be a manageable burden for his company, even if he thinks the law should have remained focused on how AI is used and not the underlying technology. So how is this being received? Well, obviously the lawmakers themselves are hailing it. Eu Commission President Ursula von der Leyen writes, I welcome the European Parliament vote on the AI Act, Europe's pioneering framework for innovative AI with clear guardrails. This will benefit Europe's fantastic pool of talent and set a blueprint for trustworthy AI throughout the world. Lawmaker Thierry Britton wrote, Democracy 1 lobby 0. I welcome the overwhelming support from European Parliament for our AI act, the world's first comprehensive binding rules for trusted AI. Europe is now a global standard setter in AI. We are regulating as little as possible, but as much as needed. And indeed that really seems to be the line that they've tried to hit. Or at least that's the narrative they're going for. Romanian lawmaker Dragos Tudoric writes, the AI act has pushed the development of AI in a direction where humans are in control of the technology and where the technology will help us leverage new discoveries for economic growth, societal progress and to unlock human potential. Not everyone agrees, however, some of the loudest commentary if you look at Twitter X are from those who feel that the law was too watered down and not strong enough. Daniel Loyfer tweets, EU officials boast about being global trendsetters and regulating AI, but with this law sets the lowest bar possible. The AI act is full of concessions to industry lobbying, exemptions for law enforcement and prohibitions full of loopholes, Amnesty International's Tech account writes. The AI act is adopted today. EU policymakers are hailing the AI rulebook as a global paragon for AI regulation, but the legislation fails to take basic human rights principles on board. Countries outside of the EU should learn from the bloc's failure going on. Amnesty writes the AI act offers only limited protections to impacted and marginalized people. It does not ban the reckless use and export of draconian AI technologies while also failing to ensure equal protection to migrants, refugees and asylum seekers. Similarly, it lacks proper accountability and transparency provisions, which will likely exacerbate human rights abuses. The notable thing here is that Amnesty's complaint is not actually about generative AI or big company power, it's more about the potential for abuse. Fanny Hidvagi, the Europe advocacy director at AccessNow, writes, The EU act is a failure for human rights, a victory for industry and law enforcement. AccessNow points out a similar set of issues as Amnesty, but also writes that the law, quote, creates a glaring loophole via Article 63 for developers to exempt themselves from the obligations of high risk systems. On the flip side are those often represented from an American perspective who think that this legislation is doomed to RUP urigate to an innovation backwater. In other words, to incentivize European talent that might create companies and AI applications at home to instead seek less regulated places to build. Lawyer Preston Byrne writes, the passage of the EU AI act once again creates a rift between the USA and the eu, which the USA can exploit with legal protection for AI development. Europe regulates, America innovates created the USA's multitrillion dollar Internet boom. It can do so again. Of course I think that we will have to wait and see to what extent that actually is the case. Will companies like Mistral give up on the European tech dream and simply return to the US and to Silicon Valley places like it? Or is that concern overblown? I am by no means an expert on this and I have not gone super deep on this act other than following along in the context of what we've been discussing here. Still, my general sense of legislation is that when people on both sides are pretty angry about it, that probably means it's hit the sort of uncomfortable compromise where most workable laws seek to hit. In other words, if one side or another is too happy about the rules, it may not represent the averaged out interest of lots of different groups. One of the only people I saw representing that perspective was the account from Growth to Value on Twitter who wrote the European Parliament has approved the AI Act. Everyone saying this is awful is either very biased or has not looked at it, and probably both. I think overall it's good. Social scoring, as is done in China is absolutely forbidden. AI models can do no emotion recognition at school or work. It's forbidden to take pictures of people on social media for databases for facial recognition. Existing AIs like OpenAI's ChatGPT will have to be clearer that a text or image is generated with AI more attention to copyright infringements. Three sensitivity categories from high risk to medium to low risk plus forbidden AI. Overall, I think this is quite balanced now in terms of where we go from here. The AI act should officially become law in May or June, and then the provisions start coming into effect in stages. Six months after the law officially comes into effect, countries will be required to ban those prohibited AI systems. A year later, the rules for general purpose AI systems start to apply, and two years after that, the whole AI act will then be enforceable. Notably, these rules apply not just to companies from Europe, but for anyone operating in Europe, and fines for non compliance can be up to 7% of worldwide annual turnover. From where I'm sitting, I don't expect a debate about these rules to end anytime soon, but they exist and I think reflect the beginning of a regulatory catch up period that is just starting right now. That is going to do it for today's AI breakdown. Until next time, peace.