Amplifying Human Intelligence

Human Intelligence > Artificial Intelligence

A deep dive into the architecture of Ubik and how our design philosophy amplifies human intelligence.

The Cognitive Cost

MIT is conducting an ongoing neurological study of brain activity in ChatGPT users while they use the AI chatbot to complete tasks. The study has three groups: ChatGPT allowed, search engine allowed, and brain only. The three groups are tasked with writing an essay using (or not using) the provided tools. MIT's early findings are concerning. Compared to the brain-only group, LLM-assisted writers were unable to complete writing tasks effectively or efficiently, and most didn't remember what they had previously generated or written. When the groups swapped, the ChatGPT writers struggled dramatically without their tool, while the brain-only group thrived the moment they were given one.

What does this mean? MIT's early data shows that LLM-dependent brains struggle without LLM assistance after exposure, or while using the LLM, which isn't surprising. Using ChatGPT is like binge-watching your favorite show while doom-scrolling on TikTok. Humans struggle to focus and remember information without friction between them and the content. Generative tools are novel in this way: a calculator cannot solve my problems without my knowledge of the formula, but one prompt to my AI friend ChatGPT and suddenly I've "written" a ten-thousand-word research paper that proves our universe is inside a black hole, as if I were a trained astrophysicist (I am not).

Critical thinking skills and intellectual agility crumble when we treat AI as an oracle, so we should build tools that amplify human intelligence rather than outsourcing cognition.

What is Ubik?

Ubik is the first AI Research Environment (AiRE). With context-aware agents, access to academic databases, and robust human approval, Ubik unlocks AI for citation-based workflows while helping build critical thinking skills: AI as a tool, not a crutch.

Unlike available AI agents, models, and platforms, Ubik agents can index, organize, and search through PDFs, create notes, and highlight text to the line level. This improves agentic capabilities, minimizes hallucination, and makes LLM-assisted work usable in high-level research and publishing. How does Ubik do this?

Agent orchestration. We build our agents optimized around their environments, not a specific model, making it easy to run them in the cloud, on devices, and eventually with local models.

Dynamic context engine. Our special sauce for structured knowledge bases, geared toward research workflows and knowledge generation. This helps the AI and the human understand the larger project at hand, as well as any research criterion: citation styling, topic area focus, finding new sources, the context of a project.

• Not just embeddings or semantic search!
• Custom document parsing and enhanced OCR for in-text citations and transparent AI output.
• Agentic analysis to extract, understand, and mark up documents beyond single questions.
• Like Cursor, Ubik knows your workspace. For improved accuracy, it references all files and agent-made items (notes, canvas) while prompting, using the @ symbol.

Custom eval suite. The Ubik evaluation suite for knowledge work and evidence attribution is designed to understand agentic performance as a research collaborator for high-level research. This covers tasks like:

• Searching for new sources and working with gathered documents.
• High-level writing and cross-document work, for publishing academic research.
• Synthesizing research and keeping citations correct.
• Compiling evidence for literature reviews and exploration.

Our complete dataset of examples and gold truths, showing the highest possible quality and accuracy in output by Ubik agents, will be open-sourced for other AI developers to reproduce and test.

Why Ubik?

Ubik is designed and developed from iterative real-world user feedback and product interviews with potential and current users. All tech is built in-house, and we are proud to contribute to the space while helping researchers and writers use AI in their work today.

Since inception, we have focused on using AI in non-generative ways. As researchers and recent graduates, the launch of ChatGPT left us astonished, excited, and worried about a future in which humans over-rely on generative AI tools.

We believe that high-quality AI output depends on skilled human guidance, and that the rise of over-accessible, query-response AI chatbots will reduce critical thinking and cognitive skills in early learners and professionals. If using generative AI tools before acquiring field-specific skills is broadly adopted, we expect to see declines in the quality of citation-based writing, a slowdown in scientific breakthroughs, and an overall weakening of the human-AI relationship.

So we started interviewing educators in NYC, across high school and higher education, and professional researchers around the country, to understand why AI was or wasn't working in their fields and what a successful AI tool would look like.

What were our takeaways?

1. Without consistent and trustworthy citation attribution, AI is unusable at the highest level of research and the lowest level of learning.
2. Without PDF interactivity and annotation tools, AI agents don't help humans turn information into knowledge.
3. High-level research is not disposable and requires human wisdom and creative thinking.
4. Successful AI for research must help build knowledge bases that last more than one chat session.
5. High-level researchers value accuracy over efficiency: high quality > speed.

From the lowest level of learning to the highest level of academia, AI chatbots were unreliable tools for educators, students, and professionals. Some teachers even said they rely on handwritten student essays, early in the year, to use as reliable sources of writing level standards per student, since generative AI has reduced the trust from teacher to student. If new technology moves us backward, how is it progress?

Our interviews confirmed humans need tools that position them in the driver's seat, not the passenger seat.

But what would an AI environment for research look like? We dove into our favorite tools and apps that help us produce work in any medium:

• Integrated Development Environments (IDEs), like Cursor.
• Digital Audio Workstations (DAWs), like Ableton.
• Cloud-based storage systems, like Google Drive.

A theme appeared: although these different forms of digital development software are used for various media and come packed with automation tools, they require intense human approval and oversight for high-quality output.

What does this mean? Cursor can code everything for you using high-powered AI tools and LLMs. But through these highly accessible and overpowered systems, new coders (vibe coders) often struggle to explain or recreate their projects after generation: Cursor is only effective if the human understands how to get to the final product without AI.

Similarly, cloud-based platforms like Google Drive have defined Gen Z's academic life, changing how we store, organize, and access personal files. With integrated tools like predictive writing and text generation powered by Google Gemini, students can now produce pseudo-polished writing that previously would have been crafted without AI assistance and taken a bit longer. But the end product would have been higher quality, and the written text would be sticky in the author's brain: good writing is not done by LLM chatbots, and text generation tools are only effective if a good writer is using them.

This is a new jump in possibilities. All these platforms are highly effective for skilled professionals, since these powerful AI tools can help experts translate their ideas quickly and alone. But through LLM chatbots like ChatGPT, the production pipeline is inverted: users come with ideas generated in full and edited by a human or another AI agent post-hoc.

Why is this important? Because generative tools help experts and hurt beginners. Quick solutions to complex problems, without friction or interactivity between the initial prompt and the desired output, don't build cognitive skills in beginners that experts have mastered without AI (vibe code, vibe physics, vibe engineering: vibe xyz = AI slop). Unlike beginners, experts amplify their human intelligence because they can correct AI output and understand when the model or agent is unhelpful or hallucinating. This is intellectual agility, a critical skill for effective AI use.

So, Who Uses Ubik?

Researchers use Ubik today because:

• They can quickly highlight text in PDFs, surface quotes in open-source papers, and confidently inject citations while prompting.
• They can search for relevant papers using Semantic Scholar and arXiv.
• They can quickly upload drafts, find errors, or expand ideas with relevant research.
• They can use tons of models (we got 'em all).

Writers use Ubik today because:

• They can use the best model for creative writing.
• They can upload long-format papers and confidently annotate, edit, and change work at speeds impossible with current AI or alone.
• They can inject quotes or highlight text in long-format papers.

Scientists use Ubik today because:

• Ubik is the only platform that offers agentic PDF searching and in-file annotation.
• Their work needs accurate evidence attribution and citation.
• They need one central place to save citations and quotes, manage sources, and edit writing with agents that can search for relevant research and amplify their unique human intelligence.

We are incredibly proud that our initial design philosophy, and our focus on using AI in non-generative ways to amplify human intelligence, are helping our current user base search and analyze peer-reviewed papers and their personal files with AI.

We have demoed with educators, researchers, and professionals, and they praise Ubik for its human-first design that promotes engaging with information rather than unquestioningly trusting answers from AI chatbots.

Ubik is the best way to do research with AI, and our goal is to help set a new standard benchmark for multi-hop research tasks. A local Ubik desktop app will be available soon. We would love to demo the app over a call to anyone interested, but for now, check out the live beta website!