Tuesday, April 7, 2026

OpenAI’s Mega Round Gets Weird

OpenAI’s Mega Round Gets Weird

Today’s Overview

Good morning, OpenAI’s giant new financing is the headline grabber, but the details are even stranger. Meta is lining up its first Wang-built models, while Google and Anthropic are both pushing harder on compute and agent workflows. There’s also fresh research on how models think before they explain. Let’s dive in.

Top Stories

OpenAI’s $122B Round Comes With Unusual Terms

OpenAI reportedly closed a $122 billion investment round at an $852 billion post-money valuation. The funding mix includes major participation from Amazon, Nvidia, and SoftBank, but the deal structure is drawing attention for more than just the size. Separate talks around a private equity joint venture reportedly include a guaranteed minimum return and early model access.

  • OpenAI’s latest round reportedly valued the company at $852 billion post-money and brought in a massive $122 billion.
  • Amazon, Nvidia, and SoftBank accounted for $110 billion of the investment.
  • A separate joint venture is said to include a 17.5% minimum return plus early access to OpenAI models.

Meta’s First Wang Models Are Nearing Launch

Meta is preparing to release the first AI models from Alexandr Wang’s Superintelligence team. Some versions are expected to be open source, while the largest models stay closed. The rollout points to a hybrid strategy even as earlier delays raised questions about competitiveness.

  • Meta is preparing to release the first models from Alexandr Wang’s team.
  • Some of the models are expected to be open source, while the largest will remain closed.
  • The launch follows delays after the codenamed Avocado model reportedly underperformed in benchmarks.

OpenAI Tests Image V2 In ChatGPT

OpenAI is testing its next-gen Image V2 model in ChatGPT and LM Arena in three variants. Early results suggest better UI design rendering, stronger prompt adherence, and improved compositional understanding. If the tests hold up, the model could strengthen OpenAI’s position against Google.

  • OpenAI is testing the model in three variants across ChatGPT and LM Arena.
  • Early tests show better UI design rendering.
  • The model also appears to improve prompt adherence and compositional understanding.

Research & Analysis

DeepMind Lets AI Rewrite Strategy Code

DeepMind’s AlphaEvolve can rewrite multi-agent learning algorithms and produced new CFR and PSRO variants. The reported methods use adaptive discounting, asymmetric regret updates, and dynamic meta-strategy solvers. They are said to outperform human-designed strategies in hidden-information games like poker.

  • AlphaEvolve produced new variants of CFR and PSRO.
  • The methods use adaptive discounting, asymmetric regret updates, and dynamic meta-strategy solvers.
  • The result is reported to outperform human-designed strategies in hidden-information games, including poker.

LLMs May Decide Before They Explain

Researchers report that LLMs often decide actions before generating reasoning tokens. A linear probe can decode those decisions from pre-generation activations with high accuracy. The result adds more evidence that chain-of-thought text may not fully reflect the model’s internal decision process.

  • The study says models often choose actions before reasoning tokens are generated.
  • A linear probe can decode those decisions from pre-generation activations.
  • That suggests chain-of-thought text may not fully capture the model’s internal decision process.

Simple Self-Distillation Boosts Code Output

A new study argues that simple self-distillation can improve code generation by fine-tuning models on their own outputs with standard supervised fine-tuning. The work frames this as a practical post-training direction for LLMs. It also looks at why such a simple method can work at all.

  • Simple self-distillation fine-tunes models on their own outputs.
  • The method is presented as a post-training path for improving code generation.
  • The study also explores why such a simple method can work.

Trending AI Tools

  • Google AI Edge Eloquent A free iOS dictation app that cleans up speech into polished text entirely on-device.

  • GitNexus Indexes codebases into knowledge graphs so AI agents can understand dependencies and call chains.

Quick Hits

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.