News

Harvard’s $699 startup bootcamp offers AI avatars of its instructors

Harvard Business School has put AI instructor avatars into its HBS Foundry startup bootcamp. The eight-week program costs $699, and HeyGen created the avatars. Participants can now receive AI feedback while practising pitches and board meetings. Live instructor sessions also take place every week.

Trial users pushed the program toward a guided experience

HBS tested an earlier concept that was closer to a chatbot. Trial users responded in a way that led the school to adopt a more guided AI experience. The avatars provide feedback during practice pitches and board meetings. Human instructors continue to lead weekly live sessions. The result combines automated rehearsal with scheduled teaching. (TechCrunch)

thinkidiot take: The program combines repeatable AI feedback with weekly live instruction.

Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research

British AI lab Inherent has emerged from stealth with a $50 million seed round and an agent called Faraday. Founded by DeepMind alumni, the company says Faraday can independently reproduce findings from published scientific papers without seeing the answers first. The agent runs on the 27-billion-parameter Qwen 3.6 model. Inherent compared its performance with Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5.

A small London team is betting on reproducible machine research

Faraday uses OpenAI's GPT-5.5 Codex as its coding tool. This separates the agent's underlying model from the software it calls to write code. Its task is to work through published research and recreate the reported findings. Inherent describes that ability as a possible step toward scientific innovation. The company has a dozen employees. They work together in person in King's Cross, London. (TechCrunch)

thinkidiot take: For developers, Faraday suggests that a smaller base model can gain useful capabilities by working with an existing coding agent.

Study explains why AI agents benefit from "skills" and when they fail

Researchers at Princeton University and UC San Diego tested why skills improve AI agents across 8,135 runs. They found that structured workflows drove most of the gains, rather than extra knowledge. Procedural grounding accounted for 65.7 percent of cases where a skilled agent did better. Directly supplied knowledge helped in 4.5 percent of tested cases.

Larger instruction libraries make the right procedure harder to find

The study also found that useful instructions can be used badly. In 10 percent of cases, agents applied an otherwise helpful skill mechanically or inappropriately. Retrieval became much worse as more skills were added. With a library of 5 skills, retrieval precision was 29.6 percent. At 100 skills, it fell to 3.3 percent. A larger collection therefore creates a selection problem before an agent can benefit from its instructions. (The Decoder)

thinkidiot take: Growing skill libraries require better retrieval and context-sensitive application.

World models that ignore human beliefs predict the wrong actions, new research shows

New research proposes Mental World Modeling, a framework that adds human mental states to simulations of the world. It represents beliefs, attention, goals, intentions, emotions, norms and social relationships. This lets models judge possible actions through both physical conditions and what people think or want. The work also introduces Menti-Bench, a set of 448 decision scenes.

The hard part is linking changing minds to changing environments

The paper says Sora, Genie 3, JEPA and Marble model the physical layer while leaving human mental states outside their state spaces. Its training-free reference implementation is called MENTIS. MENTIS divides processing into six steps. It scores action branches for physical plausibility, mental consistency and social appropriateness. Every stage produces a machine-readable intermediate result, allowing errors to be traced. Menti-Bench includes 320 text descriptions, 100 picture stories and 28 sound-video clips. (The Decoder)

thinkidiot take: Explicit models of beliefs and intentions may improve action prediction when physical state alone is insufficient.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-08-21.

SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback

SkillEvo is a system for improving reusable instructions that guide AI agents. It learns from problems that appear during longer conversations, not only from a single question and answer. Its goal is to keep finding useful corrections over repeated rounds while stopping the instructions from becoming inaccurate or unwieldy.

• Problem: Agent instructions are often written by people or generated once and then left unchanged. Existing improvement methods usually judge single exchanges, so they miss faults that emerge only after several follow-up questions and soon run out of useful feedback. A single overall score can reject a bad revision but cannot identify and repair its cause.
• New idea: SkillEvo uses simulated conversations to generate detailed feedback for each revision. Follow-up questions reveal new weaknesses over several turns, and each revision creates material for the next evaluation. A separate governance layer, meaning an independent quality-control process, repairs factual errors and unnecessary structural growth. This keeps the sequence of revisions moving in a useful direction.
• Simple example: It is like testing a recipe by cooking the whole meal and answering diners' follow-up questions. Each course may reveal a new flaw, while a separate editor checks that every correction stays accurate and does not make the recipe needlessly long.
• Evidence: Across six cloud-service categories, 9 production Skills, and 98 skill-reference files, SkillEvo beat self-reflection-based evolution by 23.0 points. It beat evolution based on single-turn question answering by 15.4 points.
• Limitation: The abstract reports tests only on cloud-service Skills and does not show whether the method works in other domains.
• Why it matters: Better feedback could let agent instructions improve for longer without accumulating errors or clutter.
• Paper: SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn

ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models

ForgeWM is a method for making interactive video simulations respond quickly to player controls. It turns a slower video generator into versions that can produce each update with very little processing. The aim is to preserve visual quality and control accuracy while keeping interaction responsive.

• Problem: Fast video generation can produce an update in only a few processing steps, but interactive simulations add a harder timing problem. Keyboard states and mouse movements must stay matched to compressed stretches of video during training and continued generation. Existing causal distillation does not by itself solve that alignment challenge.
• New idea: ForgeWM begins with a bidirectional generator, meaning a model that can use information across a video sequence. It adapts that model to the target game, trains it to predict causally from earlier information, and distills this behavior into faster student models. Distillation means teaching a smaller or faster model to imitate a stronger teacher, while on-policy distribution matching means training on the kinds of outputs the student itself produces. Separate students use steady budgets of 1, 2, or 4 denoising steps, and saved one-step drafts can later be refined for replay.
• Simple example: It is like drawing a quick live sketch while someone moves a game controller, then polishing the saved sketch afterward without replacing its original pose.
• Evidence: On paired Minecraft trajectories, ForgeWM led the evaluated systems in Imaging Quality, motion-profile agreement with the reference, action-sign accuracy, and mouse-control accuracy. It also had the lowest reference LPIPS. Replay refinement matched four-step reference quality and stayed roughly three times closer to the experienced trajectory than generating again from noise.
• Limitation: The abstract reports Minecraft results and says the training recipe transfers to gamepad-controlled FPS play, but it does not establish performance across a wider range of games or real-world environments.
• Why it matters: Responsive and controllable video models could make interactive generated worlds feel more immediate and faithful to player input.
• Paper: ForgeWM: Progressive Causal Training for Few-Step

Repo0: Design-Driven Zero-to-All Code Generation

Repo0 is a system for creating a complete software project from a written description. It designs the repository structure as part of the coding process instead of assuming that structure already exists. Its goal is to keep the project modular while turning requirements into tested code.

• Problem: Most coding agents begin with a repository whose architecture has already been chosen. Building an entire project from plain-language requirements also requires deciding how features, components, and dependencies should be organized. Without that work, the structure can become difficult to maintain as development proceeds.
• New idea: Repo0 keeps an explicit model of both requirements and software components. It represents them with a Dual-DAG, which combines a directed acyclic graph of requirements, a directed acyclic graph of components, and links showing how the two correspond. A directed acyclic graph is a network of one-way relationships that contains no loops. The system repeatedly adjusts component boundaries using measures of modularity, stops when the structure settles, and then uses that architecture to guide test-driven code generation.
• Simple example: It is like designing a house from a family's wishes by drawing one map of their needs and another map of the rooms, then linking each need to the room that serves it before construction begins.
• Evidence: Repo0 was evaluated on six real-world RepoCraft repositories with GPT-5 mini and DeepSeek V3.2. It achieved the highest Functionality Coverage and Pass Rate in every reported setting. Against RPG, it improved Functionality Coverage by up to 20.08 percentage points and Pass Rate by up to 29.74 percentage points.
• Limitation: The evaluation covers six repositories and two models, so the abstract does not establish how well Repo0 generalizes beyond those settings.
• Why it matters: Designing the architecture alongside the code could help agents build complete projects that work and remain modular.
• Paper: Repo0: Design-Driven Zero-to-All Code Generation

Trending AI Repositories

Ranking source: GitHub Trending.

openai/codex

Codex CLI is OpenAI's local coding agent. It is worth attention because it brings Codex directly onto a developer's computer, with options for editor and desktop use.

• What it is: It is a Rust project for local development work. It also connects to IDE use and a desktop app experience.
• What it does: Lightweight coding agent that runs in your terminal
• Who it helps: Codex serves developers who want to work locally, in supported editors such as VS Code, Cursor and Windsurf, or through the desktop app.
• Limitation: The IDE experience requires a separate installation.
• Repository: openai/codex

Wei-Shaw/sub2api

Sub2API is an open-source service for managing access to several AI subscriptions. It stands out for bringing shared access and cost splitting into one system.

• What it is: It is a Go service with a Vue interface. Its stack includes PostgreSQL and Redis, and it is ready for Docker.
• What it does: Sub2API 一站式开源中转服务，让 Claude、Openai 、Gemini、Grok订阅统一接入，支持拼车共享，更高效分摊成本，原生工具无缝使用。
• Who it helps: Sub2API is for people consolidating supported AI subscriptions so they can share access, divide costs and continue using native tools.
• Limitation: The documented stack includes PostgreSQL and Redis.
• Repository: Wei-Shaw/sub2api

n8n-io/n8n

n8n is a platform for building and deploying AI agents and automated workflows. It is worth attention because it is designed to carry work from a prototype into production.

• What it is: It is a TypeScript fair-code platform that sits between visual workflow design and programmable automation. It can run on your own infrastructure or in n8n's cloud.
• What it does: Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.
• Who it helps: Teams building AI agents and operational workflows can use a visual canvas, add custom code and connect their systems through more than 1500 integrations.
• Limitation: Self-hosting requires infrastructure to run it.
• Repository: n8n-io/n8n