For families looking beyond a chatbot answer

Put more research power behind a cancer case.

The bigger idea is a research workspace: bring a case’s records and scientific data together with the literature, then use AI agents to help investigate questions across them.

Agents can do more than write replies. In a properly set-up workflow, they can search, write and run analyses, check results, and follow leads—with people directing the investigation.

See what Sid’s team described

More capacity to investigate. Not a promise of a better medical outcome.

From a case to a checkable research lead
  1. 01 / The inputsRecords + molecular data + research

    Organized sources in a controlled workspace.

  2. 02 / The investigation loopAI agents work through the question
    SearchWrite codeRunChallenge

    Inspect the result → refine the question → repeat.
    People set the question, tools, and boundaries.

  3. 03 / The research outputLeads, sources, and what’s still unknown

    A hypothesis to examine—not a treatment verdict.

  4. 04 / The human decisionScientists and clinicians validate

    Test the analysis, assess relevance, decide next steps.

The important distinction

A research lead is an idea worth checking. A validated treatment option needs clinical evidence and assessment for the actual person. One does not automatically become the other.

01 / The real example behind this guide

Sid’s story wasn’t just about asking better questions.

It included technical work on biological data, with scientific collaborators.

Sid Sijbrandij, GitLab’s co-founder, and geneticist Jacob Stern described combining extensive testing, research, and AI during Sid’s cancer care. Jacob is a scientific collaborator, not a doctor.

Their March 18, 2026 OpenAI Forum conversation offers a concrete view of the process. Here are the parts that matter.

15:24–17:33 / A documented agent workflow

Roughly 600,000 cells.
An investigation you could inspect.

A concern about a possible blood-cell condition prompted Jacob to ask a focused question of his custom agent system. It worked across measurements of individual cells from several blood-sample time points—not one giant text prompt.

  1. Find what to look for. The agents reviewed literature and selected biological markers to examine.
  2. Write and run the analysis. They used code to investigate the measurements.
  3. Return the analysis, not just an answer. Jacob describes interactive plots, Python code, conclusions, and an analysis history.
  4. Have specialists investigate further. He did not treat the output as trustworthy out of the box. This was not a validated diagnostic test.

PANX3 / A lead with a missing bridge

A tumor signal met a public map of normal tissue.

Sid’s slides describe high levels of RNA from a gene called PANX3 in his tumor. RNA gives clues about which genes cells are using. A comparison with normal-tissue data in GTEx, a public reference, suggested a candidate to investigate.

But RNA activity is not the same as an accessible protein. The deck explicitly says the idea needs to look as promising at the protein level. Jacob separately describes using AI to reason through specialized laboratory tests and misleading results.

What this shows: a data comparison can generate a hypothesis with a clear next validation step. It does not establish that AI independently discovered the target.

Source: Sid’s slides, “Investing in PANX3…”; talk, 23:57–28:47.

B7-H3 / A lead challenged by other evidence

An attractive signal wasn’t the end of the story.

Jacob says ChatGPT flagged B7-H3 in a prepared gene-activity spreadsheet. In their wider program, the team considered this target using tissue staining and sequencing.

An experimental scan later showed unexpectedly high liver uptake. People revisited the data and proposed a change to the engineering strategy. The scan finding was not an AI prediction, and the proposed change did not establish safety.

What this shows: another kind of measurement can challenge a promising lead. The initial chat suggestion is not established as the cause of the program.

Source: talk, 14:36–15:24 and 22:20–23:57; slides, “B7-H3 as a seemingly good target…”.

PANX3 and B7-H3 are examples from this team’s specific investigations—not targets or treatments being suggested for your family member.

Firsthand account, not a clinical trial. The talk was hosted by OpenAI. Sid’s effort involved unusual resources, specialist teams, and experimental care. It does not establish that AI caused an outcome, or offer a treatment recipe to repeat.

02 / What the workspace brings together

Connect the evidence.
Don’t just collect more files.

A useful workspace keeps every claim connected to where it came from.

The clinical case

The diagnosis, treatment history, pathology reports, and relevant imaging reports. Keep original wording, dates, and uncertainty.

The biological data

Existing molecular tests and appropriate research data. DNA carries genetic instructions; RNA measurements help show which instructions cells are using. A specialist checks data quality and whether comparisons are meaningful.

The outside research

Original papers, methods, results, and trial registry records. Keep identifiers and dates so another person can inspect the same evidence.

Then investigate a defined question. An agent might look for relevant papers, write an analysis, compare its result with another method, and use a discrepancy to decide what to examine next.

Buying more testing is not automatically the next step. Ask the clinical team what is useful and the analyst what the available data can reliably answer. NCI explains the possibilities and limits of biomarker testing.

03 / Different tools, different jobs

A chat app is one doorway.
A research agent can do more of the work.

“Agent” means AI that can take a sequence of tool-using steps toward a task—not an independent scientist.

Read & synthesize

ChatGPT and deep research

Some questions can be explored in a research chat interface. Jacob describes using ChatGPT with a prepared RNA table and deep research to investigate specialized scientific questions before talking with collaborators.

Useful output: a source-linked explanation, comparison, or set of open questions—not a complete raw-data pipeline.

Talk, 14:36–15:24 and 25:31–27:47; OpenAI’s deep-research guide. Current app plans and file limits are not verified here.

Work with files & code

Codex and coding-agent workflows

Codex CLI can inspect local files, edit code, execute installed tools, search the web, and delegate focused work. A knowledgeable analyst can use those capabilities with appropriate scientific software and structured data.

Useful output: repeatable analysis code, results, and checks—not automatically correct biology.

Official Codex documentation, checked September 13, 2026. Specialist software, computing capacity, and validation still need to be supplied. Sid’s talk describes custom agents; it does not establish that he used Codex.

Set direction & judge validity

The people behind the tools

A cancer clinician assesses the medical context. A biological-data specialist checks the data and methods. A technically capable researcher sets up the agent workflow and reviews its work.

Essential output: a decision about what the evidence actually supports, and what still needs testing.

Software access is not the same as having this expertise. One person may cover several roles, but each responsibility needs an owner.

The custom-system route—and what to check technically

OpenAI’s deep-research API documentation describes web search, file search, connected sources, optional code execution, and cited outputs. Its Code Interpreter documentation covers Python execution, files, and graphs. These are building blocks for a developer—not a ready-made cancer research system or a promise about every ChatGPT account.

An expert still needs to check sample identity and dates, gene identifiers, data quality, normalization, cell mixtures, laboratory differences, relevant comparison groups, and false-positive risks. In plain terms: are we comparing the right things, in a sound way?

04 / What a useful handoff looks like

Ask for an evidence trail.
Not a miracle answer.

A good research handoff makes it possible for someone else to challenge the conclusion. It records the question, sources, uncertainty, and the next check.

For computational work, ask for the code, input-file versions, and checks as well as the chart. A polished graph is not proof of a sound analysis.

  1. Did the calculation work?Check the code and input data.
  2. Does it mean something biologically?Check the measurement and interpretation.
  3. Is it clinically relevant here?Check evidence for the actual person.

Passing the first check does not establish the other two. Trial eligibility and details also need human verification.

How to scope that work
Illustrative research handoffBased on a public account

Is the reported PANX3 difference worth specialist investigation?

An example of the format, drawn from Sid’s account. Not a new analysis, clinical report, or recommendation for your family.

Why this question arose
Sid’s presentation contrasts high PANX3 activity in tumor RNA with normal-tissue reference measurements.
Evidence available here
The original presentation, “Investing in PANX3…”, and firsthand talk, 23:57–28:47. We have not rerun the analysis or established a clinical evidence base.
What the actual analyst should attach
Sample dates and processing history, a labeled comparison plot, code, relevant original papers and the exact claims they support, contrary evidence, and review notes.
What is still missing
Reliable protein-level confirmation, appropriate normal-tissue assessment, and evidence of clinical relevance and safety. High RNA activity alone does not establish those things.
Question for the reviewers
What would accept or reject this lead, and is further investigation appropriate in this situation?

Status: hypothesis for expert review.
Not a treatment recommendation. The missing evidence stays visible.

05 / How a family could organize skilled help

Build a focused research effort.
Not an open-ended search for anything.

Start with one bounded piece of work and a clear person responsible for reviewing it.

  1. Choose the question and the reviewers.

    Ask the treating team whether a specialist research review could address a specific uncertainty. Ask whether a molecular tumor board—a group of specialists reviewing molecular findings—or a cancer-specific expert would help. Identify who can set up and audit the computational work.

    Agree on: the question, who reviews the answer, and the time available without delaying care.

  2. Set up a controlled workspace.

    A willing family member can coordinate records and track questions with the patient’s permission; they do not have to become the analyst. Inventory existing data, note gaps, agree on access, and keep unmodified originals.

    Agree on: approved tools, storage, outside connections, costs, and a way to export or delete the work.

  3. Commission a bounded investigation.

    Have the researcher explain which tasks agents will do: literature searches, data checks, analysis code, comparisons, or follow-up questions. Request a small first deliverable before expanding the scope.

    Ask to see: original sources, analysis files, contradictory evidence, failed checks, and a named human reviewer—not just a summary.

  4. Review the leads, then decide what merits more work.

    Scientists check the method and biological interpretation. Clinicians assess relevance, evidence, risks, and available clinical pathways. Some leads will be ruled out; others may justify further research or a verified trial discussion.

    Keep separate: an interesting mechanism, a research-use-only result, an investigational option, and an established treatment.

A short scope of work you can share with a researcher

An editorial starting point for a conversation—not a clinical protocol or instructions to upload records.

We want a bounded, expert-reviewed investigation of one cancer research question. Before accessing patient information, please propose the question, required expertise, existing data needed, approved workspace, access controls, cost, and timeline. Explain which tasks AI agents would perform and who would check them. Deliver source links, methods and code where used, evidence for and against each lead, missing information, and proposed validation steps. Separate research hypotheses from established clinical evidence. Do not select treatment or imply a result is clinically validated. Identify the clinician or scientific expert responsible for each next review.

These stages are this guide’s suggested way to structure an effort. Ask for relevant experience, scope, fees, and conflicts of interest. Specialist interpretation and data generation can cost far more than software. Some leads will have no appropriate or accessible next step; that does not mean the family failed. NCI’s guide to finding care and second opinions is one route to qualified clinical input.

06 / Boundaries that make the work more useful

Protect the person.
Make the research checkable.

Private data need a deliberate setup.

The patient’s permission, restricted access, and approved storage come before connecting tools. Check training use, retention, deletion, and external-tool access. Removing names is not enough: dates, metadata, and rare clinical details can identify someone.

Local code execution does not mean private or offline AI: relevant context may go to a hosted model. Separate public-web research from private-data analysis where practical. Codex permissions and security; HHS on consumer-app privacy limits.

More output still needs independent checking.

AI can invent citations, make analytical errors, and miss contrary evidence. Open the sources and have a qualified person review the methods. Another AI agreeing is not clinical validation. Never change, stop, or delay treatment based on an agent’s output.

WHO on AI reliability and governance. Research is not urgent care: follow the team’s urgent-contact plan; call local emergency services for an emergency.

07 / Go to the original work

Sid’s presentation, conversation, and research trail.

Personal accounts show a process. They do not establish clinical effectiveness.

Start here / original hub

Sid’s cancer page

His own starting point for the story, with links to the presentation and research resources.

Sid Sijbrandij · Living page; date not established.
Watch / firsthand process

Sid and Jacob’s OpenAI Forum talk

Start around 14:36–17:33 for data, agents, analysis, and checking; 25:31–27:47 for scientific deep research.

March 18, 2026 · Vendor-hosted talk and transcript.
Read / original presentation

Sid Cancer Journey slides

The presentation linked from Sid’s hub. Personal story with technical material—not a treatment playbook.

Living deck; read September 13, 2026. Text presentation view also available.
More from Sid’s team, with scope and access context
  • “I’m going Founder Mode on my cancer” — Sid’s January 20, 2026 (UTC) account of the project and collaborators. Personal advocacy, not clinical guidance.
  • “Assembling Your Own Tumor Board” — June 17, 2026 (UTC). His approach to gathering expertise; he acknowledges that private teams can be costly and uninsured.
  • Future of Cancer Care Today — Sid’s foundation describes its resources as free and non-clinical. Outside suppliers may charge; directory listings are not endorsements or proof of benefit, and some disclose investment relationships.

Sources checked September 13, 2026. This independent educational guide is not affiliated with or endorsed by Sid Sijbrandij, OpenAI, or a care provider. Source dates are not a medical-review certification.

The ambition is real. So is the need for judgment.

More questions investigated.
More evidence someone can check.

That is the research effort to aim for—not a promise that software can solve a loved one’s cancer.

Return to building an effort