OpenAI Forced To Delay GPT-5

In partnership with

Good morning. It’s Monday, April 7th.

On this day in tech history: In 1927, Bell Telephone Laboratories conducted the first successful long-distance demonstration of television transmission. They transmitted live video and audio of then-Secretary of Commerce Herbert Hoover from Washington, D.C., to an audience in New York City.

In today’s email:

  • OpenAI Delays GPT-5

  • Meta Releases Llama 4

  • DeepSeek’s New AI Reasoning Method

  • New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Atla has expanded what you can do with Selene 1, its evaluation-focused LLM. Originally designed to score and critique AI outputs, Selene can now be used to systematically improve them.

Developers are already seeing measurable gains. A financial research company improved output quality by 15% after integrating Selene into its workflow. A legal consultancy saw a 7% lift in answer accuracy using domain-specific evals.

Selene works behind the scenes—identifying hallucinations, flagging inconsistencies, and enforcing custom standards tailored to your use case. Improvements are automated and continuous.

Thank you for supporting our sponsors!

Today’s trending AI news stories

OpenAI Forced to Delay GPT-5 Launch: 'It’s Harder Than We Thought'

OpenAI CEO Sam Altman has delayed the launch of GPT-5 by several months, citing integration challenges and unexpectedly high performance gains uncovered during development. Instead, OpenAI will release the o3 and o4-mini models in the coming weeks—previously planned as internal components of GPT-5. The o3 model, in particular, has been internally benchmarked at the level of a top-tier coder.

Altman pointed to three key reasons for the shift: the complexity of merging features into a single system, the need to scale infrastructure for unprecedented demand, and the opportunity to push GPT-5 far beyond initial expectations. The o-series will serve as interim steps, offering scalable architecture, multimodal capabilities, and cost-effective inference.

OpenAI is preparing to release its Deep Research feature to free ChatGPT users, expanding access beyond current Plus, Teams, Enterprise, and EDU subscribers. Confirmed by Isa Fulford, a Member of Technical Staff at OpenAI, the rollout is in testing and expected to arrive “very soon,” though no specific launch date has been provided.

In a recent legal development, a federal judge denied OpenAI's motion to dismiss, calling its argument a "straw man" and upholding the NYT’s contributory infringement claim. The court noted evidence that OpenAI knew its models could reproduce copyrighted content. A new study also suggests that OpenAI’s models, including GPT-4 and GPT-3.5, may have "memorized" copyrighted material like books and news articles during training. OpenAI maintains its training practices align with fair use, but the ruling strengthens the NYT's legal challenge.

On the business front, OpenAI reportedly discussed buying Jony Ive and Sam Altman’s AI device startup according to The Information. Read more.

Meta introduces Llama 4 with two new AI models available now, and two more on the way

Image: Meta

Meta has launched two models from its Llama 4 family: Scout and Maverick, both now integrated into Meta AI across WhatsApp, Messenger, and Instagram.

Maverick, designed for general assistant tasks, contains 400 billion total parameters but activates only 17 billion across 128 experts. Scout, optimized for summarization and code reasoning, supports a 10 million token context window and can run on a single Nvidia H100 GPU.

Meta also previewed Llama 4 Behemoth, still in training, with 288 billion active parameters, positioning it as a future top-tier base model and reportedly outperforms models like GPT-4.5 and Claude 3.7 Sonnet on STEM benchmarks. However, none of these models qualify as “reasoning” models that fact-check answers.

CEO Mark Zuckerberg added that a fourth model, Llama 4 Reasoning, will be introduced within weeks. All models employ a mixture-of-experts (MoE) architecture to boost efficiency and multimodal performance. Despite their open-weight release, Meta has barred EU-based developers from using the models, citing regulatory uncertainty under the EU AI Act.

The Maverick model, ranked second on LM Arena, has raised concerns due to discrepancies between its benchmarked and public versions. Researchers noted that the LM Arena version is an experimental, "chat-optimized" model, fine-tuned for conversation, while the publicly available model differs, using excessive emojis and long-winded answers. This raises serious questions about the integrity of benchmarks, suggesting some models are engineered to perform well in tests while leaving developers in the dark about their true real-world capabilities. Read more.

DeepSeek unveils new AI reasoning method as anticipation for its next-gen model rises

Image: Liu, et.al on ArXiv

DeepSeek, a Chinese AI start-up, has unveiled a new technique to enhance reasoning in large language models (LLMs) in collaboration with Tsinghua University. The method combines generative reward modelling (GRM) and self-principled critique tuning, designed to improve LLM performance on general queries. The DeepSeek-GRM models outperformed existing approaches, offering competitive results in aligning AI with human preferences.

This development comes ahead of the expected release of DeepSeek’s next-generation model, DeepSeek-R2, which is anticipated to build on the success of its R1 reasoning model. While the company has not confirmed the R2 launch, it continues to focus on R&D, recently upgrading its V3 model for better reasoning and Chinese writing capabilities. DeepSeek has also open-sourced several code repositories for developer contribution. Read more.

3 new AI-powered tools from around the web

arXiv is a free online library where researchers share pre-publication papers.

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!