HAI Daily

Kimi for Game Development, Tencent's Hyra 2.0 Research Agents and Anthropic's Amended IPO Filing

Kimi launches a game development solution that works inside Unity, Unreal, Blender and Godot, Tencent shows Hyra 2.0 for AI-run research, and Anthropic's amended S-1 reports $17 billion in quarterly revenue.

By H.AI Tools Team3 stories
Products in this issue: Kimi, Stable Diffusion, Leonardo AI, Elicit, Consensus, Claude, Excelly-AI, Perplexity
In this issue

The short version

  • Kimi's game development solution connects its models to Unity, Unreal Engine, Blender and Godot, aiming at playable prototypes within one or two days.
  • Tencent's Hyra 2.0 splits AI research work into three layers of agents, and every conclusion must be backed by code, logs or other evidence.
  • Anthropic's amended S-1 shows third-quarter revenue of about $17 billion, up roughly 1,127% from a year earlier, and a swing to a $567 million operating loss.
  • Use AI to speed up prototypes and research, but keep the checks: playtests for games, evidence for research and the filing itself for financial figures.

The November 3 issue covers AI moving deeper into specialist work. Kimi now operates professional game engines, Tencent (腾讯) has shown a system where agents run research projects, and Anthropic's latest IPO filing shows the cost of building the models behind them. Each section explains what was announced and how to apply it to your own projects.

01

Kimi's game development solution: prototypes in one or two days

What Moonshot announced

Moonshot AI (月之暗面) launched a game development and creative work solution on November 2. It combines the Kimi K3.1 Pro model with Kimi Code and Kimi Work, and connects through plugins to professional tools: Unity, Unreal Engine, Blender and Godot. The aim is to let a team test whether a game idea is fun before committing engineers and artists to it.

In Godot, Kimi built a playable 2.5D action demo with sword attacks, dodging and blocking. In Blender it can build and modify scenes. In Unity and Unreal Engine it can read existing code, change the project and check the result; in Unity it works with Tuanjie's (团结) Codely agent, which carries out operations inside the engine. Moonshot says that in one demonstration Kimi worked for 20 hours without human intervention to deliver a Unity project that could be developed further. That is a vendor demo, not an independent test. The package also covers marketing assets, community feedback analysis, localisation and content review.

A two-day prototype plan

AI is most useful in games at the start, when the question is "is this mechanic fun?" rather than "is this production-ready?". Write a one-page brief: the core action, the goal, one enemy or obstacle, and what makes a run succeed or fail. Ask the agent for the smallest playable version in an engine your team already knows, and play it yourself within the first few hours.

Spend day two on changes driven by play, not by the agent's suggestions: adjust timing, difficulty and controls. Use placeholder art throughout. If the idea survives, image tools such as Stable Diffusion or Leonardo AI can produce concept art, but check the licence for any asset you plan to ship, and have your engineers review generated code before it enters the main project. Kimi is the direct way to try this solution. Our Stable Diffusion game development case study shows how one studio fitted AI art into its pipeline.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

Tencent Hyra 2.0: agents that run research projects

How the system works

Tencent's Hunyuan (混元) team presented Hyra 2.0 on November 2, a system in which AI agents take part in building scientific models, running research workflows and developing software tools. It divides the work into three layers. The Overall agent talks to the human researchers, aligns goals and manages priorities. Mission agents each own one research question for the long term, decide which approaches to pursue and which to drop. Worker agents read code, run experiments, diagnose failures and may challenge a Mission agent's judgement.

The system keeps three kinds of memory: project status, layered agent memory and a shared knowledge base. When a project ends, it turns methods that worked into reusable skills for later tasks. Reports say every conclusion must be backed by evidence such as code and logs, with doubtful results sent to a separate reviewer. Tencent has published research results and reports rather than a service the public can use.

Borrowing the structure for your own research

You do not need Hyra to use its design. The three layers map onto a sensible way to run any research task with AI: you set the question and the standard of evidence, one long conversation keeps the plan and the decisions, and separate short sessions do the individual searches, analyses and drafts.

Keep a running log of what was tried, what failed and why, and paste the relevant part into each new session. Ask for sources for every claim and check a sample yourself. Tools such as Elicit and Consensus find and summarise academic papers, and Claude can work through the analysis and drafting. Our AI academic research workflow shows a full example.

03

Anthropic's amended S-1: $17 billion in quarterly revenue

What the filing shows

Ahead of its IPO, Anthropic filed an amended registration statement on November 2. It reports third-quarter revenue of about $16.96 billion, up roughly 1,127% from a year earlier and up from $11.56 billion in the second quarter. Quarter-on-quarter growth slowed to about 47%, from 144% in the second quarter.

The company swung from an operating profit of $974 million in the second quarter to an operating loss of $567 million, with the cost of computing capacity rising faster than revenue. The net loss was $23.29 billion, most of it a non-cash charge from revaluing convertible notes and similar instruments. Commitments for computing purchases and leases total about $895 billion. The filing still leaves the share price range blank. Anthropic launched its roadshow the following week, covered in our November 12 issue.

The same day, Marvell agreed to buy optical-networking maker Lumentum, which we cover in our Special Edition for November 1 to 7.

Reading a filing with AI, without taking its word

An S-1 runs to hundreds of pages, and an assistant can help you find what matters if you check its work. Upload the filing, or the relevant sections, and ask for a table of revenue, operating income and net income by quarter, with the page number for every figure. Then open those pages and confirm each number. Ask separately for the risk factors that mention computing costs, and read them in full.

For tool buyers, the useful questions are whether the provider's costs are rising faster than its revenue and how dependent it is on a few suppliers. Both affect future prices and usage limits. Excelly-AI can help build the comparison table, and Perplexity can find news coverage to cross-check. This is a reading exercise, not investment advice.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

What to try next

Speed up the start, keep the checks

Use AI to get a playable prototype in two days, then let real play decide what changes. Organise AI-assisted research around one owner of the question and a log of evidence. Read financial figures from the filing, not from summaries.

Explore Code Generation, Image Generation and Search Engines to build your shortlist.

How we put this issue together

The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.

Go deeper