Interconnects
·
3 months ago
Open-weight model developers are keeping pace with closed labs on established benchmarks, but closed models maintain advantages in robustness and real-world usefulness that aren't well captured by current metrics. The author expects Chinese open-weight labs to face funding difficulties later in 2026, with capability trajectory changes visible 3-9 months after funding constraints emerge. The long-term balance between open and closed models will be determined primarily by economics and business incentives rather than pure technical capability, with open models increasingly adopted for repetitive automation tasks while closed labs maintain leadership in interactive knowledge work.
Amazon Science
·
3 months ago
Amazon and Nimbus Therapeutics fine-tuned Amazon Nova LLMs to predict molecular properties in drug discovery, matching the accuracy of specialized graph neural networks. Fine-tuned Nova 2 Lite achieved 5% lower average error than baseline GNNs while matching or exceeding them on 7 of 11 molecular properties tested. The approach reduces operational complexity by replacing multiple separately-trained models with a single model that chemists can query conversationally for predictions and reasoning.
Google DeepMind
·
3 months ago
Google introduced Gemini 3.1 Flash TTS, a text-to-speech model that adds granular audio tags allowing developers to control vocal style, pace, and delivery through natural language commands. The model achieved an Elo score of 1,211 on the Artificial Analysis TTS leaderboard and supports over 70 languages with native multi-speaker dialogue. Developers can now export precise voice parameters as API code, enabling consistent character voices across projects while all generated audio includes SynthID watermarking for AI-content detection.
Sequoia
·
3 months ago
Sequoia Capital is leading Auctor's Series A funding round for a software implementation platform that uses AI to automate enterprise software deployment. The implementation consulting market spans 9 million consultants generating over $500 billion in annual labor spend across major platforms like ServiceNow and Salesforce. Auctor's AI autopilot reduces scoping work from weeks to a single meeting by consolidating requirements, decisions, and context across disconnected systems.
Hugging Face Blog
·
3 months ago
VAKRA is a benchmark for evaluating AI agents on multi-step reasoning tasks in enterprise environments, combining API interactions with document retrieval across four capability categories. The benchmark contains 5,187 test instances across 62 domains with over 8,000 locally hosted APIs, requiring agents to complete reasoning chains of 3-7 steps that combine structured and unstructured information retrieval. Evaluation is execution-centric, assessing both the correctness of final answers and the validity of the tool-call sequences used to arrive at them.
OpenAI Blog
·
3 months ago
OpenAI updated its Agents SDK to include native sandbox execution and a model-native harness for building agents that can run long-term tasks securely. The update enables agents to operate across multiple files and external tools within isolated execution environments. Developers can now deploy more complex, stateful agents without manually handling security constraints or tool integration layers.
Hugging Face Blog
·
3 months ago
HCompany released HoloTab, a Chrome extension powered by their Holo3 computer-use AI model, that automates web tasks by navigating websites and filling forms based on user instructions. The extension includes a "routines" feature that records user actions once and then replays them automatically on demand, eliminating the need for repetitive tasks like cross-referencing competitor pricing or scraping job listings. Users can now access advanced AI task automation directly from their browser without requiring technical expertise or setup.
Together AI
·
3 months ago
Researchers introduced Parcae, a stable looped language model architecture that reuses layers multiple times to improve parameter efficiency. A 770M parameter Parcae model matches the quality of a 1.3B parameter Transformer while achieving 6.3% lower validation perplexity than previous looped models on equivalent data. The work establishes scaling laws for looped models and enables efficient training of smaller models for edge deployment by trading parameter count for increased computation within the same layer stack.
