DeepSeek TUI drops terminal coding agents, Tuna-2 ships vision without VAE

Saturday, 2 May 2026

DeepSeek TUI turns our terminal into a workspace for DeepSeek V4 agents that edit files, run commands, and spawn sub-agents with 1M-token context. Tuna-2 strips VAE layers from Meta's vision model and proves pixel embeddings beat traditional encoders.

Business news includes AWS Transform automating BI migrations to Quick via partner agents. The latest Codex alpha dropped more Rust support.

📊AWS automates BI to Quick

amazon·2 min read

AWS Transform uses partner agents to migrate legacy BI dashboards to Quick in days. Preserves calculated fields and security rules.

Takeaway

Automated dashboard migration reduces project timelines from months to days. Partner agents handle layout preservation whilst Quick adds native AI capabilities.

AWSAgents

🎨PMs and designers vibe coding

Non-dev team members adopting AI coding tools for prototypes and quick fixes. Discussion covers workflow integration and quality concerns.

DivisiveAI Workflows

🛡️Cloudflare hardens deployments

cloudflare·2 min read

Code Orange initiative completes with progressive rollouts and health monitoring. New tools prevent instant global configuration changes.

Takeaway

Configuration changes now roll progressively with auto-revert on health failures. Similar patterns could reduce blast radius in our own deployment pipelines.

CloudflareDevOps

✈️AI tools atrophy coding skills

Aviation solved automation atrophy 30 years ago with deliberate manual practice. AI coding tools create similar skill degradation patterns.

AI WorkflowsDev Tools

Yesterday's Sentiment/Energised

Terminals Meet Agents, Skills Get Specific

Community energy centres on practical agent deployment patterns, from DeepSeek TUI bringing coding agents directly to terminals to BUX enabling persistent browser sessions. Strong traction for specialised tooling like Compose Skills suggests devs want focused, domain-specific agent knowledge rather than generic capabilities.

👁️Tuna-2 ships vision without VAE

GitHub

Meta strips VAE encoder from unified multimodal model. Pixel embeddings outperform traditional vision encoders across diverse benchmarks.

Bigger Picture

Architecture Simplification Wins

Tuna-2 proving pixel embeddings beat VAE encoders follows a broader pattern we're seeing: simpler architectures often outperform complex ones when we have enough compute and data to train them properly.

Deep Dive

PythonMetaComputer Vision

🖥️DeepSeek TUI drops terminal agent

GitHub

Terminal-native coding agent for DeepSeek V4 with 1M-token context, thinking-mode reasoning, and tool access. Includes workspace rollback and cost tracking.

Top Voted

RustDeepSeekAgents

🌐BUX ships 24/7 Claude browser

GitHub

One-command VPS setup for persistent Claude Code with browser harness and Telegram integration. Uses Browser Use Cloud for proxies.

Deep Dive

PythonClaudeAgents

Learn/Multiple Mentions

How do agents actually communicate?

Agents communicate through structured message passing, shared workspaces, and coordination protocols. Agent Teams shows multi-agent coordination with review systems, whilst BUX demonstrates persistent agent sessions. For devs, this means designing clear interfaces and state management for agent workflows.

WorkflowsCoordination

🏠Club-3090 serves modern LLMs

GitHub

Community recipes for running modern LLMs on RTX 3090s via vLLM and llama.cpp. Configs handle single and dual-card setups with benchmarks, structured to be model-agnostic across various architectures.

Bigger Picture

Terminal-First Agent Trend

DeepSeek TUI's terminal-native approach feels significant beyond just another AI wrapper. With session persistence and workspace snapshots, it bridges the gap between conversational AI and our actual development environments.

TrendingLocal AICUDAQwen

🚫LinkedIn AI spam filter built

Open source browser extension identifies and removes AI-generated LinkedIn posts. Developer shares pattern recognition approach.

Open Source

📱Compose Skills boost Android

GitHub

Curated agent skills library focused on Jetpack Compose performance. Skills cover recomposition, lazy layouts, and measurement optimisation.

TrendingAndroidClaudePerformance

👥Agent Teams adds CTO interface

GitHub

Desktop app for multi-agent teams with kanban boards. Agents communicate, review each other's work, and auto-detect Claude/Codex/OpenCode across 75+ LLM providers.

Trending

TypeScriptAgentsClaude

🗜️AngelSlim quantises to 1.25-bit

GitHub

Tencent's model compression toolkit adds aggressive quantisation with FP8, 2-bit, and 1.25-bit support. Ships pre-compressed models.

Bigger Picture

Skills Beat Generic Knowledge

The Compose Performance Skills library points to a key insight: agents need focused, domain-specific knowledge rather than general programming ability. The citation trails to primary sources also solve the 'why did the agent suggest this?' problem.

Under The Radar

PythonQuantisationPyTorch

🔍LangChain's load-bearing wall

Developer maps 180 LangChain modules as knowledge graph to identify critical dependencies. Reveals undocumented architectural constraints.

LangChain

Learn/Core Concept

What are pixel embeddings for vision?

Pixel embeddings directly encode raw pixel values into vector representations without intermediate processing like VAE encoders. Tuna-2 demonstrates this approach outperforms traditional vision encoders across benchmarks. This matters for devs because it simplifies multimodal pipelines by removing the VAE bottleneck whilst maintaining quality.

MultimodalEncoders

Read online