Will OpenAI pull out GPT-5 from its hat next?
Maintaining that GPT-4 “kind of sucks”, Open AI CEO Sam Altman said recently his company will release an “amazing new model” this year. This has sparked speculation that OpenAI may release GPT-5 sometime between June and August this year. Why is this a bi
What triggered the excitement?
OpenAI, the maker of ChatGPT, inadvertently published a blog post that was indexed by search engines Bing and DuckDuckGo. Readers on X and Reddit first spotted the page, which has since been deleted but the cached version hints at the release of GPT4.5 Turbo with a “knowledge cutoff ” of June 2024 (date when the AI model will stop being trained on information). This has led many to believe that OpenAI will release GPT-4.5 Turbo this summer. When asked if GPT-5 is coming this year, Altman remarked: “We will release an amazing new model this year” but “I don’t know what we’ll call it.”
GPT-5, or whatever OpenAI chooses to call its new model, is expected to exponentially enhance the multimodal capabilities of GPT-4, have a larger context window (to allow for more inputs), and predict the next token in a sequence, enabling tasks such as sentence completion and code generation essential for chatbots like ChatGPT.
4
How’s the competition out there?
Stiff. GPT-4 has competition from foundational models like Google’s Gemini, Meta’s LLaMa, and Anthropic’s Claude 3 family. While Microsoft has invested about $10 billion in OpenAI, Amazon has upped its stake in Anthropic to $4 billion. Anthropic Claude 3 family —Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus—will initially offer a 200,000-context window, much bigger than GPT-4’s 128,000. In addition, Gemini 1.0 Ultra can run up to 1 million tokens (numerical representation of words).
Released in 2023, GPT-4’s training data specifics and parameters were not disclosed but unlike GPT-3, it can accept both text and images as input and emit a text output (multimodal). But it still cannot reason, plus it hallucinates (confidently gives wrong answers), and has resulted in a host of plagiarism and copyright violation suits. Launched in 2020 with 175 billion parameters, GPT-3 was a vast improvement over previous editions, with few shot learning (learnings from only a small number of labelled training data) but concerns remained over biases, hallucinations, and contextual understanding.
What can we expect from the “amazing” new model? 5 Is it true that GPT-4 “kind of sucks”? Are we, then, a step closer to AGI?
Artificial general intelligence (AGI) will be achieved when machines become sentient or even surpass human intelligence. However, current AI technologies are no more than excellent prediction machines that can accurately respond to specific (text, image, code, etc.) prompts because they’ve been pre-trained on huge amounts of data and are continuously fine-tuned. They still lack human-like logical reasoning, cognitive and emotional abilities (like empathy). So Big Tech is busy diluting its definitions of AGI.