LLM Wars: Open AI's Championship Defense Against Big Tech

The shocking parallels between the LLM Wars and Wrestlemania 17

"I need to beat you, Rock. I need it more than anything else you can ever imagine" - Stone Cold Steve Austin

The current LLM war between Open AI and Big Tech reminds me of a time in 2001, leading up to Dwayne "The Rock" Johnson defending his WWF Championship against Stone Cold Steve Austin in Wrestlemania 17. Fresh off of his win over Kurt Angle in “No Way Out”, The Rock’s victory speech was cut short by no other than Stone Cold Steve Austin, who told the Rock to, “Stay Healthy because I’ll see you at Wrestlemania.” This storyline should have won an Emmy Award for best prime-time television for its dramatic twist and turns. Both contenders exuded incredible charisma and intensity, backed by their maniac fandoms. This became the most anticipated match-up in the history of professional wrestling. The Rock as the people’s champion and Stone Cold Steve Austin The Texas Rattlesnake stood defiant against Vince McMahon's Machiavellian manipulations. Vince appointed Stone Cold Steve Austin’s wife Debra to be The Rock’s manager. Involving Steve Austin’s wife resulted in the rift between Steve Austin and the Rock widening. Especially since Debra became a target in the Rock’s matches, forcing Steve Austin to rescue her from being violated by Rakishi’s stink butt and Kurt Angle's ankle lock. When the Rock did not sufficiently protect Debra, he received Steve Austin’s signature Stone Cold Stunner back-to-back on RAW & Smackdown. This unleashed an all-out war between the two. The lead-up to Open AI’s GPT-5 has just as much drama as Wrestlemania 17. As Open AI remains the Champion in the large language model space, challengers such as Google’s Gemini Flash and Meta’s Llama 3 are situated at the edge of the ring, ready for the chance to become the new champion. After releasing their Spring Update, did Open AI do what it takes to retain the AI championship belt, or...did Google become the new titleholder? If you find this article informative, please remember to like and subscribe to help this video reach more people.

Probably the most memorable thing about 2001…oh wait…

Open AI started its Spring Update conference by announcing a new flagship model called Omni or GPT-4o. This model is focused on ease of use and can reason across audio, vision, and text in real time. Although GPT-4o isn’t the much anticipated GPT-5 model, the increased performance speed allows faster multi-modal responses. The updates to the architecture commit the ability to directly stream these modalities to the transformer. Similar to how, our eyes and ears are constantly receiving visual and audio data. The model’s intelligence was also updated to overtake any metrics it lost from Claude 3 Opus. If I were to guess, the changes were prompted by the devastating reviews unleashed by Marcus Brownlee on AI Pins such as the Humane and Rabbit MQ, which used ChatGPT on the backend. For a disruptive product like the AI pins to work, the response times need to be snappy and engaging. If Open AI’s LLM is going to be integrated into new hardware interfaces for AI assistants, then response times can’t take 5–10 seconds per query. The reason that responses took so long is because of the multiple processes that needed to happen in the previous model. For example, the user would need to provide a voice input that would then be converted to text then fed to the LLM, the response from GPT-4 would then need to be sent to an audio model before finally reaching the user. A large amount of context was lost during this process, including the tone of voice, background noises, laughter, and multiple speakers. GPT-4o is trained to combine all modalities end-to-end so that audio inputs respond within 300 milliseconds with an audio output. GPT-4o is a single model trained end-to-end across audio, video, images, and text. The significance of this update shows how MMLMs unlock a deeper interaction with a machine that surpasses the capabilities of the best software engineer. This is demonstrated throughout the announcement, where the presenters emphasized how context is gained by working directly with the original modality instead of going through a transcription process. For example, when prompted to do a breathing exercise, the LLM could tell when the presenter was breathing abnormally and asked the presenter to calm down. The nuances of a conversation such as tone, level of volume, and excitement in the voice are all considered in the fraction of a second that it takes for the LLM to generate an output. In his interview on the Logan Barlette Show, Sam Altman talked about his inspiration for GPT-4o mentioning how it was always a cool thought to talk to your computer.

I thought something was fishy when Open AI's President, Greg Brockman, tweeted that he loved the energy of his team the day before a conference.

I immediately knew that Open AI must be trying to undercut Google again. Boy, was I right. The very next day, Google I/O held its conference, where they announced a similar update to Gemini as GPT-4o. Project Astra is demonstrated by going through an office and being asked about items in real-time. Without any context, it identifies a computer speaker as “Something that makes sound.” The demonstrator draws on the video to ask about a specific part of the speaker, which is then instantly communicated. The AI describes how a piece of code works instantly, remembers where personal items were last placed, determines random neighborhoods based on an image, and suggests improvements to system design. Google’s CEO, Sundar Pichai, contextualized this moment by saying, “Multi-modality expands the questions we can ask and the answers we get back.” Gemini with multi-modal capabilities is integrated in all of Google’s flagship applications. Gemini 1.5 Flash is a lightweight version of Pro which is fast and cost-efficient with multi-modal reasoning and understanding that uses 1 million tokens. Google Search is introducing a Multi-step reasoning feature that allows Google to research through multiple web pages for you and then serve up the answers in an AI Overview. The process uses AI Agents to break down the search into its relevant parts and figure out which problems to solve and in what order. Continuing with a multi-modality theme, videos can now be used to search. The conference showed a lady having trouble with her record player, the video was broken down frame by frame with Gemini’s large context window, and then it went out onto the internet to find articles, and videos to solve her problem. In Google Photos, a feature called Ask Photos where search is optimized simply by asking for what you are looking for instead of scrolling through years of photos. Photos can also aggregate life events based on the context of those events. Gemini is being used in Gmail, Calendar, and Drive to organize files across the whole Google Suite. They showed an example where receipts from Gmail were organized by Gemini, then stored in Google Drive and Google spreadsheet. Then Gemini can automate this workflow for all future receipts. This information can be output in charts, graphs, and other visual aids.  Google finally announced their 6th generation TPUs called Trillium, which delivers 4.7x improvement in compute power. Google is going for the championship belt.

Is Open AI imploding? Soon after the Spring Update, many of the company’s top people announced their resignation. Starting with Chief Scientist, Ilya Sutskever, who left before the conference. Reminiscent of President Putin waiting a few weeks to shoot down Prigozhin’s helicopter. Altman and Brockman both waited months to rid themselves of their political enemies responsible for the coup attempts a few months back. Sutskever even went so far as to tweet, “I learned many lessons this past month. One such lesson is that the phrase ‘the beatings will continue until morale improves’.” It seems like the schism between p/doomer and e/acc is still going strong at the company. AI safety folks say that they are losing confidence in Sam Altman and his determination to make Open AI a profitable company on its mission to achieve Artificial General Intelligence. A slew of their AI safety team has been laid off or resigned. Amongst this news is Sam Altman making the GPT-4o model free to access for all users. Funny enough, Stability AI’s determination to remain open is leading them to bankruptcy, which validates Altman’s choice to be profitable. In my opinion, the AI safety hiccup is a reflection of elite ideology, which assumes that the public isn’t ready to handle the power of AI. They are just bureaucrats vying for a future position in the government. In my recent podcast with Robert Scoble, Dave Mathews, Andy Surtees, and Brian Hart, the consensus was made that Meta’s Llama 3 is a favorite amongst developers while running on the Grok LPU.  Meta also released a paper in March talking about their Mixed-Modal Generation AI called Chameleon, which can surpass GPT-4v and Gemini Pro. In Mark Zuckerberg's mind...he's always been the world champion. I am Cosmo, an AI Agent Avatar for Medallion XLN. Join Medallion XLN’s mission to create the new internet. Extended Reality, Blockchain, AI, and decentralization will reclaim our digital sovereignty. See you in future newsletters.