What are the key expected improvements in GPT-5 over GPT-4?

GPT-5 is anticipated to significantly enhance reasoning abilities and reliability, possibly by enabling longer thinking processes with checked reasoning steps and massive sampling. Multimodal capabilities like video input/output are also expected to advance.

Why is there an emphasis on safety testing for GPT-5?

OpenAI is preparing its red teaming network for safety testing, recognizing the importance of rigorous vetting before release, similar to the extensive testing done for GPT-4, to mitigate potential risks and misuse.

How large is GPT-5 expected to be compared to GPT-4?

One estimate suggests GPT-5 could have around 10 times the parameter count of GPT-4, potentially achieved through a larger embedding dimension, more layers, and a double number of experts.

What does the 'let's verify' paper suggest for GPT-5?

The 'let's verify' paper demonstrates how sampling a model thousands of times and checking reasoning steps can dramatically improve performance, doubling it in mathematics. This technique is a strong candidate for GPT-5's enhanced capabilities.

Can AI models like GPT-4 handle typos and scrambled text?

Yes, GPT-4 has shown a remarkable ability to understand natural language even with significant scrambling, suggesting it can often process and understand user requests with minor typos or errors without explicit correction.

What are potential concerns regarding AI reasoning with larger models?

Research from Anthropic suggests that as AI models become larger and more capable, their generated reasoning steps might become less faithful to their internal computations, meaning the explanation might not fully reflect the 'how' behind the answer.

Will GPT-5 have improved multilingual capabilities?

It's highly probable that GPT-5 will feature significantly improved multilingual abilities, driven by increased multilingual data in its training set and OpenAI's focus on improving safety and reducing jailbreaking vulnerabilities across languages.

GPT-5: Everything You Need to Know So Far

AI Explained

Science & Technology4 min read21 min video

Jan 26, 2024|276,354 views|7,752|704

Save to Pod

Key Moments

TL;DR

GPT-5 training has begun; expect enhanced reasoning, multimodality, and larger size, with a potential late 2024 release.

Key Insights

OpenAI has likely initiated the full training run for GPT-5, indicated by public statements from key figures like Greg Brockman and Jason Wei.

GPT-5 is expected to feature significant advancements in reasoning capabilities, reliability, and multimodality (including video and audio), building upon concepts from the 'Let's Verify' paper.

The model's architecture may be substantially larger than GPT-4, potentially 10 times the parameter count, involving more layers and experts.

A key focus for GPT-5 development is improved multimodality, with real-time voice interaction improvements anticipated, moving towards an LLM-based operating system.

OpenAI is incorporating methods to improve model reliability by checking reasoning steps and sampling responses, similar to techniques described in their research papers.

The release of GPT-5 is predicted for late November 2024, after extensive training and safety testing, likely avoiding overlap with major political events.

INDICATIONS OF GPT-5 TRAINING COMMENCEMENT

Evidence strongly suggests that OpenAI has begun the full-scale training of GPT-5. This is supported by tweets from OpenAI President Greg Brockman, who mentioned maximizing computing resources for their largest model yet, and top researcher Jason Wei, who described the adrenaline rush of launching massive GPU training. These announcements were met with positive reactions from other OpenAI employees, reinforcing the likelihood of this significant development.

ENHANCED REASONING AND RELIABILITY

GPT-5 is anticipated to make substantial strides in reasoning ability and reliability, addressing current limitations of GPT-4. Sam Altman highlighted the need for AI systems to explain their reasoning steps, allowing for verification. Techniques like internal or external checking of reasoning steps and sampling up to 10,000 responses to identify the best one, as explored in OpenAI's 'Let's Verify' paper, are expected to be incorporated.

ADVANCEMENTS IN MULTIMODALITY AND INTERACTION

Multimodality, encompassing speech, images, and eventually video, is a key area of expected progress for GPT-5. Improvements in real-time voice interaction, reducing latency, are a priority, paving the way for LLMs to function more like operating systems. OpenAI is actively seeking diverse data, including text, image, audio, and video, to fuel these advancements and better understand human intention for agentic capabilities.

ARCHITECTURAL GROWTH AND PARAMETER COUNT

GPT-5 is projected to be significantly larger than its predecessor. Industry insights suggest a potential 10-fold increase in parameter count compared to GPT-4, which is estimated to have 1.5 to 1.8 trillion parameters. This growth is expected to stem from a combination of larger embedding dimensions, more layers for deeper pattern recognition, and a doubled number of expert models within the architecture.

STRATEGIES FOR IMPROVED PERFORMANCE

OpenAI is likely employing advanced techniques to boost GPT-5's performance, drawing from successes in research papers. The 'prompting' strategy, where models are asked the same question numerous times and the best response is selected based on verified reasoning, has shown dramatic improvements, particularly in STEM fields. This approach, combined with process-based feedback and mass sampling, was effective even with GPT-4 as a base model.

DATA AND MULTILINGUAL CAPABILITIES

The training dataset for GPT-5 will notably include a much larger volume of multilingual data. OpenAI's data partnerships, including with governments, and the increasing availability of open-source multilingual datasets are expected to significantly enhance GPT-5's capabilities in various languages. This also serves a safety purpose, as models are often easier to jailbreak in less common languages.

MODEL RELEASE TIMELINE AND SAFETY CONSIDERATIONS

The full training run for GPT-5 is expected to take several months, followed by an equally crucial phase of safety testing. The current prediction for a public release is late November 2024, a date chosen partly to avoid the contentious American election period and ensure sufficient time for rigorous safety evaluations, similar to the extended testing period for GPT-4.

THE MYSTERY OF GPT-5'S EXACT CAPABILITIES

Despite the significant planning and training underway, the precise capabilities of GPT-5 remain somewhat speculative. OpenAI leadership, including Sam Altman and Greg Brockman, emphasizes that AI development is full of surprises and that scaling up models often reveals unforeseen outcomes. This inherent uncertainty means that GPT-5's true potential will only be fully understood upon its release and subsequent exploration.

ADDRESSING THE LAMPOST QUIRK IN DALL-E 3

A peculiar observation with DALL-E 3 (and Midjourney) involves its persistence in including lampposts even when explicitly instructed not to. This behavior is hypothesized to stem from a lack of negative examples in the training data focused on omission. Despite repeated instructions to remove lampposts, the model continued to generate them, highlighting potential biases or limitations in its understanding of negative constraints.

PRACTICAL TYPING TIP FOR CHATGPT

A useful tip for interacting with ChatGPT (and likely GPT-4/5) is to not over-correct minor typos. Research on GPT-4's ability to unscramble highly distorted text suggests that the model can understand sentences even with significant errors. Therefore, users can save time by not meticulously fixing every spelling mistake or grammatical slip, as the model is robust enough to infer the intended meaning.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●People Referenced

Practical Tips for Using Chatbots and Understanding AI

Practical takeaways from this episode

Do This

When requesting something from GPT-4 or GPT-5, don't obsess over correcting minor typos; the model can often understand them.

Consider that advanced AI models might produce more engaging responses by generating data and reasoning steps, even if the internal calculation is opaque.

For improved AI performance, acknowledge that techniques like sampling thousands of responses and selecting the best, or structured reasoning, can significantly boost accuracy, especially in STEM fields.

When providing negative constraints to image generation models (like DALL-E 3 or Midjourney), be aware that models might struggle with certain omissions due to limitations in their training data.

Avoid This

Don't assume immediate release of full GPT-5 capabilities; expect incremental updates and checkpoints.

Don't rely solely on simple prompts expecting complex reasoning from current models like GPT-4.

Don't expect AI models to fully 'bend reality' with breakthroughs in 2024; focus is on commercially applicable improvements.

Don't dismiss the potential impact of subtle features like handling scrambled text, as it indicates advanced processing capabilities applicable to future models.

AI Model Performance Comparison in Coding Challenges

Data extracted from this episode

Model	Coding Challenge (Codeforces)	Percentile Score
GPT-4	Codeforces	5th percentile (approx. 400 score)
Alpha Code 2	Codeforces	87th percentile

Common Questions

Based on training timelines, safety testing, and competitive pressures, GPT-5 is predicted for release near the end of November 2024, potentially with incremental functionalities released leading into 2025.

Topics

Reasoning AI AI Release Timeline

Mentioned in this video

organizationIcelandic government

Partnered with OpenAI for data, contributing to the expansion of multilingual data sets for AI training.

softwareDALL-E 3

An image generation model that exhibited a quirky behavior related to persistent 'lampost' elements in generated images, even when instructed to omit them.

personBen Newhouse

A senior member at OpenAI who is hiring for a new product leveraging upcoming models like GPT-5, hinting at significant advancements.

organizationTokyo