Eliza Times

Industry News

• DeepSeek-V4 launches with 1.6T parameters at $1.74/M tokens: DeepSeek releases V4 Pro (1.6T total, 49B active) and V4 Flash models with 1M context, trained on 33T tokens. The pricing undercuts frontier models significantly while claiming near-parity performance on many benchmarks. link

• Cohere and Aleph Alpha merge for sovereign AI: The two companies are forming a transatlantic AI powerhouse anchored in Canada and Germany, focusing on sovereign AI solutions for governments and enterprises. link

• GPT-5.5 achieves new Ramsey number discovery: OpenAI's latest model proved a new result in Ramsey theory, marking another milestone in AI-assisted mathematical research alongside improved coding and agent capabilities. link

Tips & Techniques

• Ask agents "Are you missing any context?": When an agent isn't performing as expected, this simple meta-question often surfaces missing information that improves results significantly. link

• Use Exa Highlights to avoid subagent error compounding: Exa can function as a hallucination-free web subagent by extracting text directly from webpages in milliseconds, eliminating the accuracy issues of traditional web-scraping agents. link

• Codex Hooks now stable for workflow automation: The latest Codex update allows automated testing, linting, and validation at key lifecycle points - "success then silent, failure then feedback" keeps agents on track without manual oversight. link

New Tools & Releases

• Sakana Fugu beta launches with recursive test-time scaling: Fugu dynamically orchestrates the best combination of open and closed models for any task, and can recursively call itself to spin up corrective workflows on the fly. Free API credits available for beta testing. link

• DeepSeek-V4 ported to MLX for Apple Silicon: Community ports enable running the 1.6T parameter model on 256GB Mac Studios, with speed optimizations ongoing. The Flash variant runs on M3 Ultra. link

• Claude Code adds Uber/Uber Eats integration: Users can now browse restaurants, check fares, see ETAs, and complete trips or orders directly through Claude without leaving the conversation. link

• Pangram Chrome Extension flags AI-generated content: New browser extension automatically detects and labels AI-generated posts as you scroll social feeds, helping users identify synthetic content. link

Research & Papers

• DeepSeek uses Meta's 2021 hash routing approach: The V4 architecture implements hash-based expert routing that Meta Research developed in 2021, plus "looped transformers" concepts from the same period - showing how foundational research resurfaces in production systems. link

• Vision models don't just differ in weights - they see fundamentally different realities: Research shows that even when trained on the same data, vision and language models develop representations of reality that diverge at a conceptual level, not just parameter-level. link

• OmniMouse: One of neuroscience's largest datasets ever: Systematic study of scaling laws in brain activity across one of the most comprehensive neural recordings assembled, with implications for understanding biological vs artificial intelligence. link

--- *Curated from 500+ tweets across AI research and development communities*

---

Emerging Trends

✨ DeepSeek V4 Release (82 mentions) - NEW DeepSeek V4 (Pro and Flash variants) launched with 1.6T parameters, new DSA2 attention mechanism, MoE architecture with 384 experts, and major performance improvements. Community quickly porting to MLX for Mac.

🔥 Codex 4M Users Milestone (95 mentions) - RISING Codex (coding agent platform) hit 4 million weekly active users, growing from 3M to 4M in just two weeks. Rate limits being reset to accommodate growth, with enterprise adoption accelerating.

🔥 Qwen 3.6 Model Release (64 mentions) - RISING Alibaba released Qwen 3.6 models (27B dense and 35B MoE variants) with strong performance for coding, reasoning, and general tasks. Becoming popular choice for local inference on consumer hardware.

📊 ICLR 2026 Conference (118 mentions) - CONTINUING Major AI/ML conference taking place in Rio de Janeiro, Brazil this week with researchers presenting papers, networking events, and workshop sessions. High attendance and engagement from global research community.

📊 GPT Image 2 Release (127 mentions) - CONTINUING OpenAI launched GPT Image 2 with significantly improved image generation quality, better text rendering, and higher resolution output. Users testing it extensively for design work, mockups, and creative projects.