- One more Thing in AI
- Posts
- #22: Sam Altman Drops Bombshell Predictions on GPT-5 and Can AI disrupt the Trillion dollar Services Industry?
#22: Sam Altman Drops Bombshell Predictions on GPT-5 and Can AI disrupt the Trillion dollar Services Industry?
#22: Latest edition of One More Thing in AI Newsletter.
Renjit Philip
April 30, 2024
I don't think of work as work and play as play. It's all living.”
Edition #22: Sam Altman Drops Bombshell Predictions on GPT-5 and Can AI Disrupt the Trillion-dollar Services Industry?
Date: 30-Apr-2024
Hi reader,
Thank you for subscribing to the One More Thing in AI Newsletter.
Specially curated for Startup founders and Business Leaders like you who want to get smarter on AI in less than 5 minutes.
This edition explores Sam Altman's GPT-5 and AGI predictions, the exploding rate of new AI models, and how AI is disrupting the trillion-dollar services industry.
We spotlight a startup slashing healthcare costs with AI and an AI company landing $150M to transform corporate conversations. Plus, a simple trick to kickstart your AI research projects.
Founders: In case you want to stress test your AI startup ideas or discuss AI automation initiatives in your business:
I hope you enjoy reading this edition, and as usual, keep learning and applying AI.
Best,
Renjit Philip
In this edition:
Sam Altman Talks AGI and GPT-5 with Lex Fridman
To understand the topic, you must listen to the entire episode of the Lex Fridman podcast, and it is worth it. For those of you who do not have two hours or more to spend on it, here is a targeted summary that I created. I listened to it at 1.25x speed while cycling (over three days!).
🎥 Sora: AI for Videos and 3D Worlds
Sora is an impressive new AI that can generate videos and 3D environments.
It can handle objects hiding behind other things (occlusions) and 3D physics. However, issues remain, such as randomly giving objects extra limbs.
This suggests it doesn't fully understand the underlying 3D world model.
Sora learns by studying massive datasets of videos and images from the Internet. Holywood, get ready to be disrupted!
⌛ GPT-5 on the Horizon
GPT-5 is expected to demonstrate significant capability increases over GPT-4.
OpenAI may release it incrementally rather than in a huge leap.
Key challenges are increasing computing power, overcoming technical hurdles, and ensuring safe behavior.
Sam wants to avoid shocking the world with radical AI updates.
💻 The Insatiable Need for Computing Power
As AI models become more capable, the demand for computing resources will skyrocket.
Sam believes compute power will become a precious commodity.
He forecasts an "uprising" in the total global computing power required for AI development.
Takeaway: Energy and power limitations will be the hardest bottleneck. We might need mini-nuclear reactors or fusion power.
🔒 Privacy Concerns with AI Assistants
OpenAI experimented with AI assistants remembering past conversations.
But this raises privacy issues about the AI storing your personal data.
Users must be able to delete anything they don't want the AI to remember. Companies need to be open about how they handle user data with these AIs that have “memory”.
🧩 Q*: The Next Leap for AI Reasoning?
Q* is a secretive project at OpenAI about improving AI reasoning abilities.
Instead of just Q&A via a chat interface, it may involve more step-by-step problem-solving.
Sam hints that Q* could represent a giant leap in AI capabilities like ChatGPT did when it was first launched. However, the specific details of how Q* works remain highly confidential.
🧠 The Path to AGI
Sam avoids speculating on timelines for artificial general intelligence (AGI). Definitions of AGI vary too much to make precise predictions.
He expects very capable but not world-changing AI by the end of this decade.
For now, the existential risk of an unaligned/rogue "AGI" is not Sam's top concern.
He believes that no single entity/person should control transformative AGI-level systems when developed. Phew!
💭 Big Dreams for GPT-5
Sam hopes GPT-5 will demonstrate broad capabilities across many domains.
It is not just optimizing narrow benchmarks but general intelligence increases.
GPT-5 could significantly accelerate the rate of scientific discovery.
Programming via natural language and developing advanced robots are goals. Aspiration is for context windows of trillions of tokens eventually.
Takeaway: I expect to see GPT -5 by this year's Q3. It is going to be groundbreaking!
Too many models, too little time!
If you are finding it increasingly difficult to keep track of the new large language models and their key features, you are not alone. We compiled a table for you below with the important ones (summarized from a TechCrunch article on this topic):
Model Name | Link | Description |
|---|---|---|
LLaMa-3 | Meta’s latest “open” flagship large language model. Widely used by the community. | |
Mistral 8×22 | A “mixture of experts” model, large size, from a French company moving away from previous openness. | |
Stable Diffusion 3 Turbo | An upgraded version of SD3, with a new API from Stability, using "turbo" in the name. | |
Adobe Acrobat AI Assistant | A tool from Adobe for interacting with documents, likely integrates ChatGPT functionality. | |
Reka Core | A multimodal model from a team previously at Big AI, competitive with major models. | |
Idefics2 | An open multimodal model built on top of smaller Mistral and Google models. | |
OLMo-1.7-7B | A larger version of AI2's LLM, one of the most open models available, precursor to a future 70B model. | |
Pile-T5 | An enhanced version of the T5 model, fine-tuned on the Pile code database. | |
Cohere Compass | An embedding model focusing on multiple data types to increase use case coverage. | |
Imagine Flash | Meta’s latest image generation model, uses new methods to accelerate diffusion while maintaining quality. | |
Limitless | A personalized AI that adapts based on user interactions across multiple platforms, including a wearable. |
Reply