Generative AI News & Interesting information - Sept 2024

Loads have happened the past month and loads of interesting things

TL;DR

1. Strawberry - Breakthrough in Open AI with release of ChatGPT o1 - yay or nay? What about SearchGPT? Agents might emerge relatively in a short span of time according to Sam Altman.
2. Gen AI heading to Hectocorn and not just a Unicorn or a Decacorn? What The Economist says?
3. Tech Giants, not VC's are leading the LLM venture capital space. OpenAI, talks to raise 6.5 Billion and if it pulls the deal off, valuations will be about $150 Billion
4. Ethan Mollick's latest Blog - Scaling: The State of Play in AI. His view on Gen AI Generations and increasing Training cost and Scaling of models including OpenAI o1. Link
5. SSI - Safe SuperIntelligence Inc. by Ilya Sutskever, Daniel Gross and Daniel Levy. What does it mean for Gen ai AND THEY RAISED $1 Billion just with paper architecture - way to go! Link
6. AGI / Super Intelligence by 2027; China / US Super Intelligence Race per LeoPold Aschenbrenner - Interesting read on "Situational Awareness" - The Decade Ahead.. Link
7. If you have not missed Eric Schmidt, the Futurist on Stanford - skipping the controversies, what he thinks about AI future? What should we learn?
8. Energy used to train OpenAI's GPT-4 Model could have powered 50 American homes for a century - as per The Economist. GPT-4 to generate 100 words consumes up to 3 bottles of water — AI data centers also raise power and water bills - are the nearby residents impacted?
9. AI to be treated at the same level of Climate Change - says UN
10. Microsoft outsourcing AI strategy and signed a 20-year deal with Constellation Energy, a power company
11. From Intel to Nvidia, AI's role in the Chip Manufacturing and how Nvidia reached a Trillion-dollar Market value milestone - how Technology supported it.

Breakthrough in Open AI with release of ChatGPT o1, how is it?

Officially called o1, nicknamed as Strawberry. I personally tried my hands-on to benchmark few of the models and found ChatGPT o1 supremely well played.

Why ChatGPT o1 and why now?

On September 17, 2024, OpenAI introduced a new line of AI models that take a significant step forward in reasoning ie., thinking well before responding and problem-solving. Designed to handle more complex tasks, these models represent a leap in performance over previous generations, with applications in science, coding, and math. With this release,

• What is offered with this release? OpenAI introduced couple of models o1-mini (uses faster reasoning) and o1-preview (uses advanced reasoning) model.
• Updated rate limits: 50 queries per week for the o1-preview model and 50 queries per day for o1-mini. This move aims to balance availability and performance for early adopters.
• Enhanced Reasoning Abilities: OpenAI's new models spend more time "thinking" through problems, improving their human-like reasoning and problem-solving skills.
• Significant Performance Improvements: The models achieved 83% on an International Mathematics Olympiad qualifying exam, a substantial jump from the previous model's 13%.
• Coding Excellence: They excel in coding tasks, ranking in the 89th percentile on platforms like Codeforces, showing advanced coding and debugging capabilities.
• Rebranding and Focus: Despite lacking features like web browsing, the models are powerful for complex reasoning tasks and are rebranded as the OpenAI o1 series.
• Emphasis on Safety: OpenAI has enhanced safety measures, with the new models better at avoiding harmful outputs and collaborating with institutions to ensure responsible deployment.

Who Will Benefit from these Models?

The potential applications of OpenAI’s new models are vast, especially for those working in fields that require complex reasoning. Scientists, researchers, and developers stand to benefit the most. For example:

• Healthcare researchers can use o1 to annotate complex datasets, such as cell sequencing data.
• Physicists working on cutting-edge topics like quantum optics can use it to generate the intricate mathematical formulas their research demands.
• Developers across various fields can leverage the model’s ability to execute multi-step workflows efficiently.

What are some of the strengths & weaknesses?

Strengths:

• Excellent Reasoning Abilities: Thinks through problems very effectively, even outperforming some experts.
• Flexible Problem-Solving: Can adjust its approach when it spots mistakes, leading to better solutions.
• Improved Safety and Alignment: Understands and incorporates human values, making its decisions safer.

Weaknesses / Limitations:

• Less Effective in Simple Tasks: Not as good at tasks that don't require deep thinking, so it's more specialized.
• High Computational Demands: Requires a lot of computing power, which can make it hard to use in everyday situations.
• Feature gaps. At launch, the o1 models lack web browsing, image processing and file uploading capabilities.
• API restrictions. At launch, there are a variety of restrictions on the API limiting the models. Function calling and streaming are not supported initially. There is also limited access to chat completion parameters during the preview phase.
• Response time. OpenAI users have come to expect rapid responses with little delay. But the o1 models are initially slower than previous models due to more thorough reasoning processes.
• Rate limits. For ChatGPT Plus or Team users, OpenAI initially limited o1-preview usage to 30 messages a week, rising to 50 messages a week for 01-mini. On Sept. 16, 2024, OpenAI raised the limit for o1-preview to 50 messages a week and increased o1-mini to 50 messages per day.
• Cost. For API users OpenAI o1 is more expensive than previous models -- including GPT-4o.

What are some of the use-cases?

Sample reference: https://www.youtube.com/watch?v=5EPazPx59UU

Is next evolution SearchGPT?

https://openai.com/index/searchgpt-prototype/

With OpenAI experimenting SearchGPT,

• OpenAI's SearchGPT Prototype: OpenAI is experimenting with SearchGPT, an AI search feature that provides fast and timely answers with clear and relevant sources, functioning similarly to Google Search.
• Potential for Enterprise Search: If successful, SearchGPT could be adapted for internal enterprise searches, similar to tools like Glean.
• Impact on Marketing and SEO: Marketers and advertisers are closely monitoring SearchGPT due to its potential impact on advertising, marketing, SEO, and website attribution.
• Limited Availability: Powered by OpenAI models like GPT-3.5, GPT-4, and GPT-4o, SearchGPT is currently available as a prototype to a small group of users and publishers.
• Growing Interest: Those interested in testing SearchGPT's features are paying close attention as more details emerge.

ChatGPT 40 vs ChatGPT o1

A New Form of Scaling - Thinking

• New Approach to Scaling: OpenAI's o1-preview and o1-mini models introduced a different kind of scaling, focusing on inference compute (the AI's "thinking" time) after training, rather than just model size.
• Post-Training Thinking: Instead of traditional scaling during training, these models improve by performing multiple internal reasoning steps (hidden thinking tokens) before producing an output.
• Chain of Thought: The models follow a step-by-step reasoning process, which enhances accuracy. The longer they "think," the better the answers become, revealing a new scaling law.
• Exponential Scaling: Similar to training, this "thinking" scaling law has no limit, but the longer the AI thinks, the more computing power is required for improvement.
• Early Days: While still in its infancy, this new approach to scaling shows great promise for AI’s future capabilities.

Gen AI heading to Hectocorn and not just a Unicorn or a Decacorn?

Generative AI is reshaping Silicon Valley, pushing the creation of "hectocorns"—startups valued at $100 billion—far beyond unicorns ($1 billion) and decacorns ($10 billion). OpenAI is reportedly raising $6.5 billion, which could value the company at $150 billion, making it America's second hectocorn after SpaceX, according to The Economist.

Since 2019, Microsoft has invested $13 billion in OpenAI, while Amazon has invested $4 billion in Anthropic. Tech giants, not traditional VCs, are leading investments in large language models (LLMs). These companies also provide cloud computing power to train AI models and distribute their products—OpenAI via Microsoft Azure and Anthropic via AWS. Microsoft is expected to invest more in OpenAI's latest funding round.

Ethan Mollick's latest blog:

In the latest blog by Ethan Mollick (Link), titled "Scaling: The State of Play in AI" (Sept 16, 2024), he delves into the current capabilities of AI models, particularly Large Language Models (LLMs) like ChatGPT and Gemini. Ethan Mollick explains that AI development has largely been driven by scaling—bigger models with more parameters and computing power perform better at complex tasks. These "frontier models" rely on vast amounts of data and computing resources, leading to advancements in intelligence and performance across various fields, such as translation and financial analysis.

Ethan Mollick identifies a new scaling law that applies after a model is trained. This law relates to the model's ability to improve performance by allocating more compute power to "thinking" during inference. Essentially, this process involves generating multiple reasoning steps internally before producing an output. The longer a model "thinks," the more accurate the response, akin to how human thought improves with reflection.

With two scaling laws now evident—one for training and another for "thinking"—Mollick suggests AI capabilities will continue to grow rapidly. While it’s unclear when we’ll hit a limit in training larger models, improvements in the "thinking" process guarantee further progress in tackling complex problems. Mollick predicts that with these advancements, AI systems capable of handling sophisticated tasks with minimal oversight are just around the corner, marking a significant frontier in AI development.

This is a significant article and 2 topics piqued my interest

In Ethan Mollick’s article, he discusses two key topics: Generations of AI models and o1 scaling.

1. Generations of AI Models and each requiring ten times more computing power and data than last:

Mollick categorizes AI models into generations based on their size, capabilities, and the amount of compute power required to train them. He simplifies the classification into three generations:

• Gen1 Models (2022): These include models like ChatGPT-3.5, requiring less than 102? FLOPs of compute, with training costs under $10M. Many open-source models also fall into this category.
• Gen2 Models (2023-2024): Represented by models like GPT-4, these require between 102? and 102? FLOPs and cost over $100M to train. They have significantly more capabilities, with multiple Gen2 models now available.
• Gen3 Models (2025-2026): These models, like the upcoming GPT-5 and Grok 3, are expected to require between 102? and 102? FLOPs and could cost billions to train. Gen3 represents the next leap in AI development.
• Gen4 Models and Beyond: Expected in the coming years, Gen4 models could exceed $10B in training costs, and scaling could continue up to 1,000 times beyond Gen3 by the decade’s end.

Large Language Models (LLMs) require immense energy. For example, training GPT-4 used enough electricity to power 50 homes for 100 years, and generating just 100 words consumes up to 3 bottles of water for cooling, as per OpenAI CEO Sam Altman. Training costs for the largest models now reach $100 million, with future models possibly hitting $1 billion or more. Even generating responses, such as summarizing financial reports, can cost between $2,400 and $223,000 per query. In response to the energy surge, Microsoft signed a 20-year deal with Constellation Energy to expand AI and cloud technologies, with nuclear energy emerging as a reliable, carbon-free option to meet uninterrupted power demands.

2. o1 Scaling:

• Mollick highlights OpenAI's o1-preview and o1-mini models, which introduce a new approach to scaling, different from traditional size-based scaling. The innovation here is in the post-training inference compute, or how much computational power the model uses during inference (or "thinking").
• The o1 models improve performance by generating multiple reasoning steps (or "thinking tokens") before producing a final output. This approach to “thinking” scaling shows that the longer the model spends "thinking" about a problem, the better its answer becomes, following its own scaling law. Mollick notes that this process mirrors the exponential scaling law of training, where greater compute leads to better results.
• This new form of scaling suggests that even after we hit limits in training model sizes, improvements in inference compute (i.e., allowing the model to think longer) could continue to enhance AI performance, expanding the model's capacity to solve complex problems.

As Mollick writes, “AI capabilities are poised for dramatic improvements in the coming years... this dual-pronged approach virtually guarantees the race for more powerful AI will continue unabated, with far-reaching implications for society, the economy, and the environment.”

Ilya Sutskever and LeoPold Aschenbrenner - Amending the Future of AI:

Safe Superintelligence: A New Player in AI

Safe Superintelligence (SSI), co-founded by OpenAI's former chief scientist Ilya Sutskever, recently raised $1 billion to develop AI systems that surpass human capabilities, with a focus on safety. Valued at $5 billion, SSI aims to build a trusted team of top researchers and engineers. Investors include major venture capital firms like Andreessen Horowitz, Sequoia Capital, and DST Global.

The company’s mission is to create AI that aligns with human values, in contrast to other industry players like OpenAI, which has faced internal conflicts over leadership and safety governance. Sutskever’s departure from OpenAI after a board disagreement, and his dismantled “Superalignment” team, marks a shift in his focus. Now, SSI is aiming to carve its own path, focusing on safety and scalability with a more conventional corporate structure.

Impact on OpenAI:

OpenAI could face increased competition, particularly around AI safety—a key concern in the industry. With Sutskever, a key figure in AI development, launching SSI, OpenAI will need to continue innovating to maintain its position. Moreover, SSI’s commitment to safety and alignment may draw investor and regulatory attention, challenging OpenAI’s dominance in these areas. However, OpenAI’s established scale and partnerships still provide it a significant edge in both infrastructure and funding, but competition is clearly heating up.

LeoPold Aschenbrenner, Situational Awareness - AGI by 2027?:

In the Article (Link), Key Points from "The Decade Ahead" by Leopold Aschenbrenner

1. From GPT-4 to AGI (Artificial General Intelligence)

• AI advancements are rapidly accelerating. The leap from GPT-2 to GPT-4 was significant, evolving from preschool-level intelligence to high-school-level in just four years. By 2027, AGI—systems that can perform any intellectual task a human can—seems highly plausible. This progression is driven by advancements in compute power and algorithmic efficiency.

2. From AGI to Superintelligence

• Once AGI is achieved, progress will not stop at human-level intelligence. A rapid transition from AGI to superintelligence—AI systems vastly smarter than humans—will occur. The acceleration will be so fast that a decade of advancements could happen within a year.

3. Racing to the Trillion-Dollar Cluster

• The AI race is triggering an extraordinary industrial and economic shift. By the end of the decade, trillions of dollars will be invested in building datacenters, GPUs, and power infrastructure to support AI development. US electricity production is projected to grow substantially to meet the demands of this new AI economy.

4. Locking Down AI Labs for Security

• Current AI labs are neglecting security, risking critical AGI knowledge falling into the wrong hands, particularly state actors like China. As AGI progresses, there will be a massive need for securing AI models and systems from external threats.

5. Superalignment Challenge

• Controlling AI systems that are much smarter than humans remains an unsolved problem. Ensuring that superintelligent systems align with human values is critical, and failure in this could lead to catastrophic consequences during the intelligence explosion.

6. Global Competition: The Free World vs. Authoritarian Powers

• Superintelligence will provide significant economic and military advantages. The race for AGI puts the survival of democratic nations at stake, as authoritarian powers like China are still major players. Maintaining leadership and avoiding self-destruction is essential for the free world.

7. Government Involvement: The Project

• As the AGI race heats up, governments, especially the US, will become deeply involved. By 2027-2028, the US is expected to establish some form of government-backed AGI project. This involvement is necessary because startups alone cannot manage superintelligence development responsibly.

What It Means for the Future:

• Rapid AI Advancements: AI is moving at lightning speed. By 2027, systems will not only match human intelligence but surpass it shortly after. This rapid progress means leaders must prepare for the significant societal, economic, and security implications of superintelligence.
• Industrial Transformation: The AI race will drive an unprecedented industrial shift, with trillions invested in infrastructure and energy production. This will change how industries operate, and leaders need to think about future-proofing their strategies.
• Security Risks: AI labs are currently vulnerable to breaches. As AGI nears, securing AI from malicious actors becomes a top priority for leadership across industries and governments.
• Alignment Challenge: The biggest technical challenge is ensuring that superintelligent AI systems act in line with human values. If not managed well, the risks are existential.
• Global Competition: The race for AI dominance is also geopolitical. The free world must stay ahead of authoritarian regimes, like China, to maintain economic and military power.
• Government’s Role: No single company can handle superintelligence alone. Governments will play a crucial role in regulating and supporting AI development to ensure safety and fairness. Leaders in both public and private sectors must collaborate to guide this powerful technology responsibly.

Eric Schmidt, futurist view on Future of AI

Few key takeaways for me

1. ??AI outsourcing is all not that bad - Microsoft has outsourced their most important AI milestone of the Century to OpenAI head by Sam Altman and going strong. Microsoft is one of the strongest.
2. ??How Nvidia has reached a trillion-dollar market capitalization, making it the first semiconductor company to achieve this milestone? Which key technical factor provided this ability?

• Focus on AI and GPU Technology: Nvidia pioneered the development of GPUs (Graphics Processing Units), which have become essential for AI, machine learning, and data processing. Their chips are now the backbone of AI infrastructure, powering everything from cloud computing to autonomous vehicles.
• Early Investment in AI: Nvidia recognized the potential of AI early on and invested heavily in R&D for AI-specific chips. This foresight allowed them to dominate the market as demand for AI applications skyrocketed.
• Strong Ecosystem: Nvidia didn’t just sell chips—it built an ecosystem. Its CUDA (Compute Unified Device Architecture) platform allowed developers to optimize software to run on Nvidia GPUs, creating a lock-in effect and boosting adoption across industries.
• Strategic Partnerships: Nvidia’s collaborations with major cloud providers like Amazon, Microsoft, and Google have solidified its position in the AI and cloud computing space, expanding its reach and customer base.
• Continuous Innovation: Nvidia continuously pushes the boundaries of technology, from GPUs to AI chips like the H100, keeping it ahead of the competition.

3. AI Advancements and Impact

• ??Large context windows and text-to-action systems will have an impact bigger than social media in the next year, enabling AI to process million-word prompts and trigger arbitrary digital commands.
• ??LLM agents are being developed to autonomously read, discover principles, test, and expand understanding in fields like chemistry, potentially revolutionizing task automation.
• ??The combination of context window expansion, agents, and text-to-action will have unimaginable impacts on AI systems, accelerating fields like chemistry and material science.

4. Global AI Competition

• ????The US is 2-3 years ahead of China in AI development, with a 10-year chip advantage due to the ban on Nvidia chips in China.
• ??The AI industry is experiencing a potential investment bubble, with mind-boggling amounts of money being invested and big investors adding AI components to everything.

5. AI Development Challenges

• ?Non-Transformer architectures like state models and long context models are limited by memory and CPU/GPU speeds, with supercomputers dominating the field.
• ???AI development favors the rich due to the need for huge capital and technical expertise, while countries with limited resources must find partners to compete.

6. Entrepreneurship in the AI Era

• ?The ability to prototype quickly using AI tools is crucial for entrepreneurs, as competitors can now build similar prototypes in days, intensifying the challenge of innovation.

AI to be treated at the same level of Climate Change - says United Nations

• ?? Global AI Oversight: The UN proposes an international body to monitor and regulate AI, similar to climate change oversight.
• ?? Collaboration and Standards: Nations should cooperate to share data, set AI standards, and help developing countries engage in AI governance.
• ?? Addressing AI Risks: Concerns include job loss, misinformation (like deepfakes), and social bias, all of which could be worsened by unchecked AI growth.
• ??? Human Rights Focus: AI governance should be based on human rights to ensure protection and fairness globally.
• ???????? Competition Among Nations: The US and China, along with other powers, have competing views on AI regulation, making global consensus difficult.
• ?? Need for Coordination: A hybrid approach of both global and national governance is recommended to manage AI development effectively.
• ? Urgency: Experts stress the need for immediate action as AI technology advances rapidly beyond current regulatory capabilities.