This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.
If you follow AI for work, research, investing, or just to understand where the technology is heading, this weekly brief is a concise way to scan the most important launches, risks, and resources in a few focused minutes.
The week of 16th to 22nd August 2026 brought a strong mix of multimodal model releases, agent infrastructure, developer tools, and practical AI adoption. RedNote released dots3-note preview, an open-weight 280B MoE model built for long-horizon multimodal agents, while DeepSeek introduced V4-Flash-Vision-Exp for visual agent workloads and OpenRouter debuted the mysterious, high-performing Ox Alpha coding model. On the product side, Cartesia shipped Sonic-3.6 voice AI, Wispr raised a $280M Series B, and Meta expanded its AI offering with a native Mac app and business-oriented integrations.
The week also highlighted the rapidly maturing ecosystem around AI operators: Artificial Analysis launched a benchmark for agent web-search APIs; K-Dense released drug-discovery agent skills; Harvey opened a legal-engineering certification; and Replit introduced a credit-free mode for routine agent work. Google made AI plans free for eligible students across more than 140 countries, while Bun 1.4 delivered a major Rust-based runtime overhaul for JavaScript and TypeScript developers. Alongside the releases, Anthropic CEO Dario Amodei’s public defence of AI-risk discussions and his optimism on AI-enabled biomedical progress underscored the continuing tension between capability, adoption, and safety.
The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interest
Xiaohongshu/RedNote announced dots3-note preview, an open-weight multimodal MoE AI model from RedNote's AI lab aimed at advancing long-horizon agency for real-life tasks like reasoning, environment exploration, memory updates, and tool use.
Key specs include 280B total parameters with 16B active, 512K context window, native support for text/vision/video/audio, and TEMPO—a new RL method using self-critiquing and test-time scaling for extended agent training.
Accompanying benchmark charts show competitive performance against larger models on agentic, coding, and multimodal tasks; model weights are on Hugging Face, with API access now available via OpenRouter.
Anthropic CEO Defends AI Risk Talk While Eyeing Disease Cures
In a detailed X exchange, Anthropic CEO Dario Amodei pushed back on criticism from investor Gavin Baker that his risk warnings fuel anti-AI views. Amodei stressed that real breakthroughs, like curing most human diseases in 5-10 years, will outweigh messaging, and revealed Anthropic's push into biology and medicine with early results expected soon. He defended regulation proposals that aid smaller players while tackling cyber, bio, and alignment risks, as prediction markets buzz with trader bets on his timeline
Voice AI startup Wispr raised a $280M Series B led by Menlo Ventures at a $2B valuation, taking its total funding to $361M, and now has 100K business customers
Harvey, an AI company specializing in complex legal work, announced the launch of its free self-paced Legal Engineering Certification via Harvey Academy.
The course teaches four core skills—legal expertise, AI fluency, process design, and user enablement—to help professionals design AI-enabled legal workflows and drive adoption.
Aimed at lawyers, in-house teams, and legal tech roles, it includes practical training from over 1,000 AI implementations and awards a shareable badge upon completion.
Cartesia announces Sonic-3.6 TTS with major gains in naturalness across 44 languages, including realistic pauses, filler words in English, and fluid Hindi-English code-switching, released three months after Sonic-3.5 based on user feedback.
Sonic-3.6 ranks #1 on Artificial Analysis's Provider Voice leaderboard (Elo 1282) and Controlled Voice leaderboard (Elo 1123), ahead of Speechify Simba 3.2, Alibaba's Qwen-Audio, and ElevenLabs v3 in blind tests using both native and standardized voices.
Pricing: At $49 per 1M characters via the Cartesia platform, Sonic 3.6 is the most expensive of the top three models on quality, roughly 5x Speechify AI's Simba 3.2 at $10 and nearly double Alibaba's Qwen-Audio-3.0-TTS-Plus at $27.59, though still half the price of ElevenLabs' Eleven v3 at $100
Available immediately in beta via cartesia.ai, the state-space model architecture delivers low-latency streaming performance, positioning Cartesia as a leader in production-grade, human-like voice AI.
Artificial Analysis Launches Web Search API Benchmark for AI Agents
Artificial Analysis released its Search Index, evaluating web search APIs like Parallel, Exa, and Firecrawl on real agent tasks such as research and multi-hop browsing. Parallel leads with a 75 score, followed by Exa at 74 and Firecrawl at 73—all far ahead of the model-only baseline of 33. The benchmark factors in total cost and time, highlighting how top performers like Parallel and Firecrawl balance efficiency with savings on model tokens. Developers praise the independent data for picking reliable tools as AI agents lean more on web search.
FWIW, I wrote this public Notion page on the benefits of integrating external search/fetch providers into a harness
K-Dense AI announced the release of a new agent skills plugin specialized in small molecule and protein therapeutics, enabling AI agents to handle tasks in cheminformatics, molecular modeling, docking, dynamics, and target discovery.
The plugin is a curated subset of their main Scientific Agent Skills GitHub repo (K-Dense-AI/scientific-agent-skills), which has 33.8k stars, 161+ validated skills, and 100+ scientific databases covering biology, chemistry, medicine, and drug discovery
It supports installation and use in multiple AI coding tools including Claude Code, Cursor, Codex, and Google Antigravity via the open Agent Skills standard.
Replit's new "Free Mode" lets Core and Pro subscribers use the Agent for everyday tasks (chatting, ideating, quick builds) without consuming usage credits, effectively letting them create up to 30X more within their existing monthly subscription.
Free Mode is powered by OpenAI's GPT-5.6 Luna and applies to fast, high-value everyday tasks — chats, ideation, quick edits, and simpler builds — which no longer draw down credit-based usage.
Just LDA asked the following question
Anyone here using Buzz, Grok Bot, or the new Hermes Bot function?
Opinions? I’d assume Hermes Bot works better than Grok bot because then you’re not limited to one model.
But just curious to hear from others that have used them.
I’m currently using Buzz and it seems to be working well. I have my Claude agent, Hermes agent (ChatGPT), and Grok agent all working together now on my various projects.
Diamond Hands Dig responded to Just LDA with the following
Just wrote about Hermes Agent Bot Mode yesterday
Google made AI Pro free for US students, AI Plus free for students in 140+ countries, and a cheap YouTube student bundle.
Google Gemini announced a back-to-school promotion offering eligible US college students one year of Google AI Pro for free and students in 140+ countries one year of Google AI Plus at no cost, running through December 31, 2026.
New student tools rolling out include a dedicated hub for study notebooks, flashcards, and quizzes; adaptive lessons that pinpoint knowledge gaps from class materials; and interactive 3D visualizations generated via prompts like “Show me how DNA works.”
Gemini Live now supports hands-free multi-step research reports that run in the background, with notifications and voice discussion of findings to support deeper academic work.
Whilst not LLM related per-se, thought this might be of interest to group members particularly those who work with Javascript/Typescript codebases which are coded/generated via their coding agents and maybe utilising nodejs or deno for their runtime
Bun 1.4 released (Claude Code has been running on a canary version of it for quite a while)
The rewrite of Bun from Zig to Rust in 11 days was done with the assistance of Fable 5
PS: Anthropic acquired Bun last year. Jarred Sumner is the founder
Couple of links on the Meta AI app (which doesn't seem to have rolled out to my region, might need to VPN or Meta could have region locked based on metadata from my usage of FB/IG). The video in Spencer Barnett's post makes it very intriguing to try it out
DeepSeek Launches Experimental Vision Model Matching Top Benchmarks
The new DeepSeek-V4-Flash-Vision-Exp adds image-handling to the V4-Flash series, scoring big on multimodal benchmarks like 36.5 on ApexBench and 27.3 on Agents' Last Exam—often rivaling Opus-4.8—while holding steady on pure-text tasks. Developers access it easily via the DeepSeek API with base64 images, URLs, or the fresh Files API, all at standard pricing. Community voices, from researchers to tech analysts, praise the leap in visual agent performance without speed or cost hits, signaling a push toward versatile multimodal tools.
Mysterious Ox Alpha AI Model Debuts with Strong Coding Prowess
Ox Alpha, listed on OpenRouter as a free preview from an anonymous provider, handles 1 million token contexts with text, image, and video inputs, excelling in coding and agentic tasks. Early tests by developer Ben Davis showed it hitting 80% on DeepSWE benchmarks, outpacing models like Fable at 65% and gpt-5.6-sol at 52%. Community speculation points to Chinese origins, possibly Zhipu’s next GLM or Xiaomi’s MiMo, with its polished speed and frontend strengths fueling excitement during the limited free access period.
Below is my personal website which aggregates links to many of my socials as well as the various content and community that I curate. Feel free to share this link to others who you think may find this content/community useful to them
The cover image of this newsletter via generated via the Seedream 5 Pro model within the Krea tool via the following prompt
An enormous floating city in the sky, steampunk style with lots of towers and domes, oil painting, fantasy art, detailed


280B MoE with 16B active parameters, a 512K context window, and multimodal understanding across text, vision, and audio








on GitHub).









New stealth model: Ox Alpha
100T tokens per day free capacity, no 5hr usage limits.
