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 1st to 7th November 2026 highlighted AI’s move towards more complete workflows, with agents taking on the coordination and follow-through needed to turn a request into a usable result. DeepSeek’s dsh-superagent research prototype organised development requests into parallel tasks for human review, while Cursor’s Cloud Agents brought app creation, databases, hosting and security checks into one workflow, leaving publication to the user. The significance lies in how much work these systems can carry between human checkpoints. Better models help, but so do the tools that organise tasks, test results and make the finished work usable.
The cost and trust questions become more consequential as those workflows expand. Revolut’s account of self-hosted, open-weight coding models suggested that cheaper, less restricted access can broaden how teams use AI, while Together’s compiler work targeted the ongoing cost of running changing workloads. Yet wider access is not an automatic outcome: Bloomberg reported Goldman Sachs restricting advanced coding tools for most staff after reviewing productivity and risk concerns. My updated Perplexity Decisions guide explored a related issue as web search became available for structured judgements: current information can improve a decision, but a model’s confidence does not establish that its sources support it. Completing more steps only helps if the result is worth its cost and can withstand scrutiny.
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
DeepSeek Releases Proactive Coding Agent Prototype
DeepSeek unveiled dsh-superagent, an open-source research prototype that monitors GitHub and optionally Feishu for requests, creates isolated task tickets, and completes work in parallel sessions before seeking human approval. In tests with 300 messages and 20 real requests, it cut median task times to 12.4 minutes, far faster than single-session baselines. Developers are already adapting it with their own prompts and harnesses, praising how structured scaffolding boosts agent performance without needing new models.
An agent framework that works through a repo's GitHub issues and PRs with many agents in parallel, designed to take as little human attention as possible while delivering as much value as possible.
github.com/deepseek-ai/ds…
Revolut Switches Engineers to Cheap Chinese AI Model for Huge Gains
via a talk from Ops Lead David Abitbol in AI Engineer NYC
Europe's biggest private company made a Chinese open-weight model the default for its 2,000 engineers, running it on 56 Nvidia GPUs at just 30 cents per million tokens—about one-tenth the cost of tools like Claude or Codex. Lead times dropped 70 percent and coding time fell 60 percent after the April 2026 switch, as unlimited access let teams tackle Jira tickets, docs, and workflows beyond just code. Frontier models stay available for edge cases, while tuning handled rising demand without extra hardware, signaling a shift toward open-weight economics for large teams.
The first chapter of modern fintech was powered by smartphones, cloud computing and easier licensing. That phase is largely over. The next one is being shaped partly by artificial intelligence — specifically by
Shipping like a startup, regulated like a bank | AI Engineer NYC 2026
David Abitbol describes Revolut's move toward self-hosted coding models, the inference tuning that made them practical, and the operational controls used around a regulated engineering organization.
Anthropic confirms new Claude Code limits “to steer it away from recreating someone else's specific game”, saying IP infringement is against its usage policy
Original games, genre games, and games built from your own designs and assets are meant to
Last week some of you ran into Claude Code pushing back on game projects it should have just built. We'd added a short rule to the Opus 5.5 prompt to steer Claude away from recreating specific existing games, but it was also kicking
Original games, genre games, and games built from your own designs and assets are meant to
PS: If you have done any reasonable amount of webdev then Addy Osmani is a name you've very likely encountered
Goldman Sachs has irked some of its staff after it banned most of them from using advanced artificial intelligence tools designed for coding after reviewing productivity and risk concerns
Unfortunately the main Bloomberg article is behind a paywall so I have also linked to some tweets which provide commentary. I assume they have read the paywalled article and other relevant material
: Michael Nagle/Bloomberg
- The bank has limited access to Anthropic’s Claude Code and Microsoft’s GitHub coding tools to engineers, following internal guidance updated in October.
- Traders and bankers had previously used them to monitor
The approach used several model providers at once: an in-house assistant built on models from OpenAI, Anthropic and Alphabet, plus Anthropic's Claude Code and
Nvidia Directs Data Centre Partners to Cut China Exposure
According to the Financial Times, Nvidia has told its data centre and cloud computing partners to ensure at least 51% of their services are sold to non‑Chinese customers, as part of reducing exposure to Chinese clients
Nvidia orders data centre clients to reduce China exposure
World’s biggest chipmaker tells cloud computing providers a majority of their clients must be non-Chinese
via Yat Siu
thoughts on this? Basically saying chinese models are compromised on US service...real or propaganda?
What's In America's Code?
Can the SI models writing and powering our nation's code be trusted? Here's what we found when we put LLMs to the test.
Tencent Launches Full Hy4 AI Model with Vision and Computer Control
Tencent's Hy4 a 770B-parameter model adds native vision for detailed image analysis, math from visuals, and computer-use tools to handle screens, mouse, keyboard, and apps like Blender and CAD, positioning it among leading models for productivity in finance, coding, games, office tasks, and research.
Benchmark charts show Hy4 excelling in areas like program synthesis, agent exams, automation, browser tasks, and OS interaction, often matching or surpassing competitors such as DeepSeek, Kimi, Claude, and GPT variants.
The release highlights responsible development focused on cyber and biosecurity
Available now on Tencent Cloud and OpenRouter at $0.834 per million input tokens and $2.501 per million output tokens, with model weights coming soon and a lighter Flash version in the works. Currently there is a 50% discount via OpenRouter
Hy4 is here, new with vision.Firmly among the leading open models.
More productive everywhere: finance, frontend, game, officework, research…
Built responsibly, with cyber/biosecurity in mind.
Try it out and let us know!
：hy.tencent.ai/research/hy4?l…
Across +8.5K real-world, long-horizon agentic sessions from our global community of
Hy4 is here, new with vision.Firmly among the leading open models.
More productive everywhere: finance, frontend, game, officework, research…
Built responsibly, with cyber/biosecurity in mind.
Try it out and let us know!
：hy.tencent.ai/research/hy4?l…
Hy4 - API Pricing & Benchmarks
The official release of Hy4 delivers comprehensive and substantial improvements over Hy4 Preview, with particularly strong gains in agentic capabilities and...
OpenAI launches Images 2.6 in ChatGPT, Codex, and the API, saying it generates images 50% faster, and adds animated GIF creation from prompts or photos
Image-2.6 Flare and Sunburst, our two new API image models, generate images 50% faster than before.
At medium or high quality, Sunburst keeps details more consistent from one edit to the next.
Introducing Images 2.6
Images 2.6 brings faster image generation, GIF creation, more precise editing, and new API controls across ChatGPT, Codex, and the API.
I’ve updated the Perplexity Decisions guide and companion skill I shared earlier.
Perplexity’s Decisions API can now optionally search the web before returning a judgment—useful for questions about changing facts, such as whether a newer software release is available.
The guide explains how this works, how it compares with Parallel, and what changes in costs and evidence handling. The key point: a confident answer doesn’t necessarily mean the sources support it.
The skill now covers these details too. If you installed an earlier version, replace the complete folder.
Chen-Ping Yu, whose AR navigation startup Phiar was acquired by Google, announces Atmee v2: a real-time video foundation model for lifelike AI avatars that generate any character from one image plus 10 seconds of audio, enabling contextual expressions, gestures, and environment interaction.
Atmee v2 supports live video calls, voice-only, and video generation modes, claiming up to 50x faster inference and 10-100x lower cost than models like Seedance 2.5 and Wan 3.0, optimized specifically for character-driven content over general scenes.
The launch thread showcases practical demos including unscripted live avatar talk shows, mobile app integrations for tutors and DJs, 3D game characters you can converse with, and early experiments toward agentic living worlds, with access via signup at atmee.ai.
After two years of research, today we’re launching Atmee v2: a
Simran Arora announces the Together Compiler (TKCC), an agent-driven system for continual AI inference optimization that adapts to rapidly shifting "polymorphic" traffic patterns where 80% of OpenRouter usage involves models from the last three months.
TKCC achieves 252x higher optimization experiment throughput via a Tiered Persistent Engine for hot-reloading candidates, a parallelism cost model accounting for data-dependent behaviors, and targeted agent abstractions, delivering SOTA tokens-per-second results like 1205 TPS on DeepSeek v4.1 on limited hardware in 24 hours.
The work builds on Together's prior ThunderKittens and Megakernels research, now serving trillions of production tokens daily with quick recompilation; the team shares a detailed paper and blog to advance open research on agentic compilers for production-scale inference.
Today’s inference is “Polymorphic” or shape-shifting, it doesn’t sit still. On OpenRouter, 80% of traffic goes to models released in
via Juan Rivieria Perez
Just launched the best way to use AI to automate your social media growth. Would love to hear what you think if you use AI to help run your social media.
Cursor now lets you build and share live web apps directly from a prompt.
Cloud Agents handle the full stack: scaffold the app, set up Postgres if needed, configure Cursor SSO auth, provision hosting, run security checks, and deploy to a unique shareable link. Nothing goes live until you explicitly publish.
Use cases already happening inside Cursor:
- Internal ops tools (desk seating, event registration)
- Sales dashboards
- Onboarding experiences
- Customer presentations
- Engineering portals
Available on paid plans with Cloud Agents. Update Cursor and try it.
Links
- Tweet/demo: https://x.com/cursor_ai/status/2118458629091426304
- Changelog: https://cursor.com/changelog/build-live-apps
- Docs: https://cursor.com/docs/cloud-agent/apps
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 ChatGPT Image 2.6 model within the Krea tool via the following prompt
watercolor an animated portrayal of daily life