Jason Says
Watch Qwen3.8-Max dropping open-source this week — Sonnet-class performance at $2/M tokens signals the open-vs-closed capability gap is closing fast; August may be the inflection point for indie devs to rethink their model stack.
OpenAI's GPT-Live introduces a turnless speech model and low-latency architecture for continuous voice interaction — no waiting for pauses. Built over six months, it's a foundational shift for voice AI developers building real-time conversational agents.
Qwen3.8-Max, a 2.4T-parameter model, dethrones Kimi K3 as the open-source leader with Sonnet-class performance expected. Set to be fully open-sourced this week at competitive pricing ($2/M input, $6/M output), it's a significant milestone for the open-source AI ecosystem.
AWS now allows vibe-coding tool Superblocks to be embedded inside AWS customers' private clouds — a major enterprise distribution milestone. The bigger implication: AI apps are decoupling from their underlying models, enabling secure on-prem AI development at scale.
Apple's AI overhaul finally makes Siri a capable assistant — but the moment feels flat. In a landscape where competent AI assistants are table stakes, arriving late means the upgrade lands with a thud rather than a bang. A cautionary tale for slow-moving incumbents.
Baseten's $13B Series F signals inference engineering is now a standalone discipline. The Latent Space masterclass covers everything from autoregressive to diffusion model serving — KV cache, batching strategies, latency-throughput tradeoffs — essential reading for AI infrastructure builders.
Anthropic's constitutional midtraining inserts a values-based training stage between pre- and post-training at 120B scale. Unlike shallow RLHF alignment that erodes under fine-tuning, this approach produces durable alignment — critical for enterprise deployments where model robustness under adaptation matters.
Current RLVR methods like GRPO uniformly spread reward across all tokens — ignoring that some tokens matter far more to the outcome. This paper proposes counterfactual sensitivity credit reallocation to identify and reward high-impact tokens, meaningfully improving long-CoT reasoning training for developers doing RLVR fine-tuning.
EMBL's AI Librarian wraps Europe PMC's 40M+ life-science records into an agent-native knowledge layer. Instead of forcing AI agents to learn complex search syntax, it provides structured, directly callable interfaces — a strong vertical-domain template for MCP/RAG tool builders.
Investors: Marc Benioff 参投
用 AI 解决 AI 落地难题的初创公司，今日从隐身模式曝光，获 Salesforce 创始人背书，聚焦简化企业 AI 部署流程。
为 AI 模型引入「审美判断力」的人类评估平台，拥有 530 万用户，为前沿实验室提供设计领域的关键人类反馈数据。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free