Anthropic is preparing Claude Opus 4.7 for release, along with a new vibe coding tool, according to The Information.
LLMS
-
Google’s Internal Code Models Outperform Claude
By
–
Perhaps it depends on which tech director he was chatting with. Google's internal coding models are better than Claude, because they are trained on Google's own codebase, and they are being used extensively. But there may be large pockets of lameness within Google?
-
Anthropic Updates Claude Code with Epitaxy Feature Set
By
–
Anthropic is rolling out "Epitaxy", an upgraded version of its Claude Code solution for desktop.
— 🚨 AI News | TestingCatalog (@testingcatalog) 14 avril 2026
> Simultaneous work on multiple repositories
> Updated layout with terminal, plan, tasks, and preview sections
> Loads of hotkeys and more
Did you get it already? 👀 https://t.co/9NqJKTcivC pic.twitter.com/U4kfbzxQRqAnthropic is rolling out "Epitaxy", an upgraded version of its Claude Code solution for desktop. > Simultaneous work on multiple repositories
> Updated layout with terminal, plan, tasks, and preview sections
> Loads of hotkeys and more Did you get it already? -
Essential Reading on Bash and Agents
By
–
who's written about bash re: agents? what's essential reading? (putting something together for the newsletter)
-

Coinbase Shipped AI Agent Services with LangGraph and LangSmith
By
–
@Coinbase Developer Platform serves over 10,000 individual developers building on crypto APIs. At Interrupt, the Agent Conference by LangChain, Evan Kormos (Engineering Manager, Coinbase) will walk through how his team shipped AI agent services with LangGraph and LangSmith,
-

Nexus: Improving LLM Generalization Beyond Pretraining Loss
By
–
Fascinating underexplored scaling direction "Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima" This paper claims that pretraining loss is not a sufficient proxy for downstream capability. Since LLMs trained on mixed domains can reach the same
-
TinyFish AI: 87% Token Reduction, 2x Task Completion
By
–
4/ The workflow also feels way cleaner. The output goes to the filesystem instead of just filling up your context window, and TinyFish says it uses 87% fewer tokens than MCP per operation, with 2x higher task completion on more complex multi-step tasks.
-
Qwen 3.5’s overthinking helps infer intent, boosting Gemma 4’s performance
By
–
Qwen 3.5 tends to overthink, which ironically helps it better infer intent Gemma 4 is the opposite, you have to spell everything out (probably guardrails) Had Qwen rewrite my prompts, then used those on both models Gemma’s performance jumped noticeably
-
Reinforcement Learning Speed Trade-offs in Large Language Models
By
–
Significant downside of larger models is that it is much harder to do RL on them. Smaller model = quicker RL cycles, bigger model = slower RL cycles. So far this matches – GPT-5 is the smallest model with quickest iterations, Gemini 3 is biggest and slowest iterations, Claude is
