ChatGPT $100 Pro Plan Introduced (2 minute read)
OpenAI introduced a new $100/month ChatGPT Pro tier aimed at power users, expanding its pricing lineup between the $20 Plus and $200 Pro plans. The company confirmed the $200 plan still exists despite not being listed on the pricing page.
|
Making Claude Cowork ready for enterprise (3 minute read)
Claude Cowork is now enterprise-ready, enhancing work with organization controls like role-based access, group spend limits, and expanded observability features. Admins can manage adoption with detailed usage analytics and integrate tools like Zoom for seamless workflows. Companies like Zapier and Airtree have already leveraged these features for improved project management and operational efficiency.
|
Perplexity launches Personal Finance powered by Plaid (2 minute read)
Perplexity has expanded its financial services with Plaid integration, transitioning from portfolio tracking to a comprehensive finance dashboard. Users can link checking, savings, credit card, and loan accounts to analyze spending, track liabilities, and calculate net worth. Targeted at AI-savvy consumers, it combines data insights with Plaid's infrastructure, transforming from a portfolio add-on to a full personal finance hub without executing trades.
|
|
Agentic Infrastructure on Vercel (8 minute read)
Vercel argued that coding agents were reshaping how software gets built and deployed, with agent-initiated deployments rising sharply and accounting for more than 30% of weekly deployments. The post framed this shift as a need for new infrastructure designed for agents to deploy software, run AI systems, and increasingly operate infrastructure autonomously.
|
CoreWeave Takes As Much Financial Engineering As It Does Datacenter Design (4 minute read)
CoreWeave's revenue backlog is up to $87.8 billion, with Meta represents 40.1% of that backlog. OpenAI's deals represent 25.5% of extended backlog. CoreWeave only has two dozen named customers, but it likely has more. The company achieved $5.13 billion in sales in 2025, up 2.7x from 2024, and posted a net loss of $1.17 billion. It recently held a private offering of $1.75 billion in notes to cover its data center expansion costs.
|
|
Anthropic launches advisor tool for Claude Platform API users (1 minute read)
The new advisor tool on Anthropic's Claude Platform API provides developers with the ability to use Opus as an advisor alongside Sonnet or Haiku as executors. This allows agents to access advanced reasoning capabilities while maintaining operational costs at the level of the more efficient executor models. The advisor tool is now publicly available through a configuration in the Messages API request.
|
Multimodal Embedding & Reranker Models with Sentence Transformers (8 minute read)
The v5.4 update of Sentence Transformers introduces multimodal embedding and reranker models that enable encoding and comparing texts, images, audio, and videos within a shared embedding space for tasks like cross-modal search and retrieval-augmented generation. These advanced models map different modalities into a consistent embedding space and assess relevance across mixed-modality pairs, though lower cross-modal similarity scores may occur due to modality clustering. Notably, multimodal rerankers provide high-quality scoring by evaluating individual pairs for more accurate retrieval results.
|
Process Driven Image Generation (22 minute read)
Meta has introduced a multi-step framework for image synthesis that alternates between textual reasoning and visual updates. This enables models to iteratively plan, draft, reflect, and refine outputs.
|
Efficient Rollout Scaling for Diffusion RL (4 minute read)
NVIDIA's Sol-RL introduces a two-stage framework that separates exploration and training, using FP4 rollouts to generate large candidate sets and BF16 for selective policy updates. This approach reduced compute costs while improving alignment and accelerating convergence in diffusion model post-training.
|
|
Dario Says Continual Learning Is Solved. Is It? (5 minute read)
Continual learning research is aimed at breaking past the feasible horizon of current techniques. If models can't learn new things while performing tasks, they will struggle when the task horizon grows very long. Context management and other engineering improvements can push task horizons to weeks, or even months, but very long-horizon tasks remain out of reach. Human-level continual learning may be 'solved', but the revolution is not yet complete.
|
Introducing KellyBench (2 minute read)
KellyBench evaluates sequential decision-making in sports betting by simulating the 2023-24 English Premier League season. Models like Claude Opus 4.6 and GPT-5.4 struggled, with none achieving positive returns, highlighting their limitations in adapting to long-term strategies. This underscores a need for complex environments enabling agents to learn from experience under uncertainty.
|
|
|
Love TLDR? Tell your friends and get rewards!
|
|
Share your referral link below with friends to get free TLDR swag!
|
|
|
|
|
Track your referrals here.
|
|
|
|
