AI Demos: Big Drops This Week from OpenAI & Google | E1950

18 May 2024 (7 months ago)
AI Demos: Big Drops This Week from OpenAI & Google | E1950

Sunny joins Jason to dive into this week’s AI news and demos. (0s)

  • Sunny joins Jason to discuss the latest AI news and demos.
  • OpenAI released several new demos this week, including:
    • A desktop app that can take screenshots and photos.
    • An image captioning tool that can generate detailed descriptions of images.
    • A text summarization tool that can condense long articles into shorter summaries.
    • A language translation tool that can translate text between different languages.
  • Google also released several new AI demos this week, including:
    • A music generation tool that can create new songs in different styles.
    • A text-to-image generator that can create images from text descriptions.
    • A chatbot that can answer questions and generate text.

OpenAI's Big Reveal on Monday (2m10s)

  • Sam Altman appeared on the All-In podcast but didn't mention OpenAI's upcoming demos.
  • The demos were pushed back, so Sam Altman couldn't discuss them on the podcast.
  • Some people criticized Sam Altman for not sharing more information on the podcast.
  • Future versions of ChatGPT may not be numbered.
  • Sam Altman shared the story of his firing from OpenAI for the first time.
  • OpenAI is working on chips and a phone, which were previously undisclosed projects.
  • Sam Altman discussed copyright and training data, revealing that OpenAI won't touch music due to copyright concerns.

A dive into what OpenAI launched this week. (5m40s)

  • OpenAI released several new features this week, including improved voice control, math problem-solving assistance, and screen-reading capabilities.

Improved voice control (5m40s)

  • OpenAI's voice control now allows users to adjust the speed of the AI's responses, making it more natural and user-friendly.
  • The AI can now be interrupted and will adjust its responses accordingly.
  • OpenAI's AI can now tutor users on math problems, providing guidance and nudges in the right direction without giving away the answer.
  • This feature is particularly useful for students who want to improve their understanding of math concepts.
  • OpenAI's AI can now read and understand the content on a user's screen, allowing it to provide assistance with tasks such as solving math problems or navigating websites.
  • This feature has the potential to make the AI more versatile and helpful in a wider range of situations.

LinkedIn Jobs - Post your first job for free (10m11s)

  • LinkedIn Jobs is a platform that helps businesses find qualified professionals for their teams.
  • It has over a billion members in more than 200 countries, making it a great place to post job openings.
  • 70% of LinkedIn users don't visit other leading job sites, so posting on LinkedIn ensures a wider reach.
  • 86% of small businesses using LinkedIn Jobs get a qualified candidate within 24 hours.
  • Posting the first job on LinkedIn Jobs is free.
  • There is now a language model that can provide adaptive learning, which is customized to the user's learning level and pace.
  • This technology can be used to learn various skills, such as playing chess, math, cooking, or playing guitar.
  • The language model can provide personalized guidance and feedback, similar to a human tutor.
  • This technology has the potential to provide billions of people with access to customized tutoring.

OpenAI’s new model (13m12s)

  • The new model can reason across audio, vision, and text in real time.
  • It can now start a live video stream and describe what it sees.
  • The model can be used in "environmental awareness mode" to provide context-aware assistance.
  • It can understand the context of a scene and provide relevant information or suggestions.
  • It can listen to sound and provide context based on what it hears.
  • The new capabilities include voice conversation and showing the AI what the user is talking about.
  • The new features provide a more intuitive and user-friendly interface.

Open AI’s “memory” setting and their new MacOS desktop app! (17m21s)

  • Added a "memory" setting in ChatGPT that allows users to customize responses based on previous interactions and preferences.
  • Users can ask the AI what it remembers about them, and it will provide information based on past conversations.
  • Launched a Mac desktop app that includes features such as taking screenshots and providing image analysis.
  • The AI can analyze and describe the content of images, including identifying objects, emotions, and activities.
  • The desktop app currently lacks a continuous monitoring mode, but it may be added in future updates.
  • Google introduced Bard, an AI chatbot designed to provide information and generate text.
  • Bard is powered by PaLM 2 (Pathways Language Model 2), a large language model from Google AI.
  • Bard aims to combine the breadth of knowledge from the web with the power of Google's language models.
  • It can generate text, answer questions, and provide information on a wide range of topics.
  • Bard is still in the testing phase and is not yet widely available to the public.

Eight Sleep - Head to and use code TWIST to get $350 off the Pod 4 Ultra. (19m43s)

  • Eight Sleep's Pod 4 Ultra can cool down each side of the bed to 20° F below room temperature.
  • It also has an adjustable base that fits between the mattress and bed frame, allowing for different positions for reading and sleeping.
  • The Pod 4 Ultra tracks sleep time, sleep phases, HRV, and heart rate.
  • The user can get an extra hour of deep sleep every night because of it.
  • The user invested in the company because they love the product so much.
  • The user imagines a virtual podcast producer that could research, write blog posts, post to LinkedIn, and draft emails.
  • The virtual assistant could also summarize the day's accomplishments and remind the user of forgotten tasks.
  • Such persistent assistants exist in task management software, but listening to the user's entire day would provide a whole new level of power.

Jason highlights the startup Athena. (22m37s)

  • Jason invested in a company called Athena, which provides real human assistance based in the Philippines for $3,000 per month.
  • Athena's assistants are highly skilled and efficient, and they can collaborate with users on various tasks.
  • Jason believes that the combination of Athena's assistance and AI tools like My Assistant Plus can greatly enhance productivity.
  • The persistent app feature that monitors the user's activity and provides real-time assistance is a game-changer in AI.
  • Jason rates the contextual awareness and environment watching features of the app as A+, but gives the ability to talk to it better a B.
  • No relevant information to summarize.

Looking at the improvements on the new ChatGPT 4o model. (25m53s)

  • ChatGPT 4 is faster and has low latency.
  • ChatGPT 4 is truly multimodal, trained on text, video, and audio simultaneously.
  • The "O" in Omni stands for Omni, representing audio, visual, and text processing in real time.
  • Voice conversations are now available on the desktop app for Plus subscribers.
  • Costs for businesses are expected to decrease, with a goal of making the service free for all users in the future.
  • ChatGPT 4 is currently not free, but there might be a path to free access in the future.
  • OpenAI may adopt a model where the latest version is paid while previous versions become free over time.
  • Omni is expected to be expensive, while ChatGPT 3.5 is relatively cheap.
  • The cost difference between ChatGPT 4 and 3.5 is not specified, but ChatGPT 4 is surprisingly cheaper, costing half as much.

HiddenLayer - HiddenLayer’s AI Detection & Response Solution (29m31s)

  • HiddenLayer provides a security solution for generative AI and large language models (LLMs).
  • It offers visibility and tools to detect and prevent threats like confidential information leaks, malicious prompt injection, and model theft.
  • Trusted by leaders in finance, tech, healthcare, and even the US Department of Defense.
  • Enables companies to harness the power of AI securely and confidently.
  • Discussion about the recent AI demos from OpenAI and Google.
  • OpenAI's ChatGPT is a text-based AI that can generate human-like text, write code, and even write poetry.
  • Google's Imagen is a text-to-image AI that can generate realistic images from text descriptions.
  • Both AI demos showcase the rapid advancements in generative AI technology.
  • Discussion about the future trajectory of AI models.
  • The argument that models will not keep getting bigger due to increasing costs and diminishing returns.
  • Tech usually gets more efficient, not bigger and more expensive.
  • Future models may be cheaper to run than older models, rendering older models obsolete.
  • The rapid depreciation of older AI models as newer, more efficient models are released.
  • Example from Croc: Llama 2 was deprecated within 48 hours of Llama 3's release.
  • Older models become less cost-effective compared to newer models with better efficiency.

Breaking down the pricing of the new vs historic ChatGPT models. (32m34s)

  • The pricing of the new ChatGPT models is significantly lower than previous versions.
  • GPT-4 is 50% cheaper than GPT-3, with a cost of $5 per million tokens input and $15 per million tokens output.
  • Open-source models are even cheaper, with a cost of 60 cents per million tokens input and 80 cents per million tokens output.
  • The original GPT-3 model was priced at $120 per million tokens input and $60 per million tokens output, which has dropped to $15 per million tokens input and $5 per million tokens output over the past year.
  • It is predicted that the cost of these models will continue to decrease significantly over time, becoming 100 times cheaper every two years.
  • The lower pricing of AI models makes it more feasible to use them for a wider range of applications, including continuous recording and processing of audio, screenshots, and video.
  • This will lead to a significant increase in the amount of data processed by AI models.

Jason demos Recall. (35m31s)

  • OpenAI's "Recall" feature summarizes and saves YouTube videos to a knowledge graph, allowing users to easily access and organize information. It automatically builds connections between different topics and concepts, creating a personalized knowledge graph for each user.
  • Google's generative experience will be integrated into its search engine, Gmail, YouTube, and Chrome, creating a comprehensive knowledge graph of user activities.
  • OpenAI and Google have recently made significant advancements in AI technology, with OpenAI's language model knowing a lot about users, including their interests and preferences, which raises privacy concerns.
  • Google Chrome now has built-in AI features, such as "Help Me Write" and the ability to create custom images and organize tabs intelligently.
  • Developers can build apps that are significantly better than existing ones by continuously adding features and listening to customer feedback.
  • Google has expanded the context of its AI model, Gemini, to 2 million, launched a new AI chip, and made more services available within Vertex.

Sunny shares [Labs.Google]( and NotebookLM. (42m19s)

  • Google launched NotebookLM, a tool that can take up to 1.5 million tokens of input and generate summaries.
  • NotebookLM can be used as a research assistant to compile information and provide key moments and lessons learned from the input.
  • The tool has the potential to replace researchers in various fields, as it can perform basic research tasks efficiently.
  • Chat GPT has been used by producers to summarize content, and its capabilities are comparable to human summaries.
  • With 1.5 million tokens of input, NotebookLM can cover a significant amount of published content, including blog posts and podcast transcripts.

Sunny shares [Labs.Google]( and NotebookLM. (42m19s)

  • Google introduced a generative search experience, which is a preferred search method for Sunny.
  • The generative search experience provides more detailed and comprehensive answers to search queries.
  • It can generate text, code, and other types of content based on the user's input.
  • The generative search experience is powered by large language models and aims to provide users with more relevant and useful information.

[Labs.Google]( and “AI Overviews” (46m3s)

  • Google's AI Overview feature provides instant and detailed responses to user queries, potentially impacting SEO and the traditional flow of traffic from Google search to external websites.
  • OpenAI and Google recently released impressive AI demos, with OpenAI being less cautious about potential lawsuits due to its new player status.
  • A startup is creating a clearinghouse for the rights of content holders in AI, similar to Storyful for viral videos.
  • Retweeting or sharing content on social media is generally safe, but downloading and monetizing copyrighted content without permission can lead to legal issues.
  • Characters in the United States are protected by copyright for the author's life plus 70 years, with the "Mickey Mouse rule" extending the term to 95 years for works published before 1978.
  • This week in startups will have a new co-host, Alex Wilhelm, joining Sunny and Jason, aiming to produce 250 episodes per year, with 150 hosted by Jason and 100 by guests.
  • The speaker is considering changing the name of the "Weekend Startups" segment to "Twist" and prefers a six-day workweek over a seven-day workweek.

Overwhelmed by Endless Content?