ChatGPT conversations with engineers routinely produce useful practical knowledge that is currently lost after the conversation ends.
I propose an explicit opt-in “Contribute to Community Knowledge” feature that would allow a user to select part of their own conversation and turn it into a structured, publicly searchable engineering note.
Nothing would ever be contributed automatically. The user would explicitly select the material, review the resulting note, confirm the right to publish it, and authorize the contribution.
AI would organize and maintain the knowledge; engineers would provide, review, reproduce, challenge, and correct it. Contributions could contain calculations, simulations, measurements, datasheets and references, successful solutions — and, importantly, documented failures.
The goal is not another forum or encyclopedia, but a Collective Engineering Laboratory Notebook: a versioned repository of reproducible engineering knowledge recording what worked, what failed, under which conditions, and why.
The problem
Engineers accumulate an enormous amount of practical knowledge that rarely makes it into textbooks, datasheets, application notes, or academic papers.
Much of it consists of apparently small observations:
- a component that works unexpectedly well in a particular application;
- a circuit technique and the conditions under which it works;
- an undocumented interaction between components;
- a simulation revealing a failure mechanism;
- an unsuccessful experiment showing why an apparently reasonable solution does not work;
- a practical workaround discovered after several iterations;
- an obsolete component whose unusual properties make it particularly suitable for a niche application;
- a modern replacement that appears equivalent but fails because of some secondary parameter.
Individually, many of these findings are too small to justify a formal publication.
Collectively, they represent an enormous body of engineering knowledge.
Today much of this knowledge disappears into private notebooks, individual ChatGPT conversations, old forum threads, laboratory computers, or simply engineers’ memories.
ChatGPT creates an unusual opportunity to preserve it because much of the reasoning, experimentation, correction, and interpretation already happens inside the conversation.
Proposed feature
Allow the user to deliberately select a useful fragment of a ChatGPT conversation and choose:
Contribute to Community Knowledge
Nothing should ever be contributed automatically.
ChatGPT would transform the selected material into a structured draft containing, where applicable:
- Problem / objective
- Application context
- Proposed solution
- Circuit or system configuration
- Components involved
- Experimental or simulation conditions
- Results
- Negative results
- Failure modes
- Root-cause analysis
- Applicability limits
- Datasheets and primary references
- Measurements, simulations, schematics, photographs, or other evidence
- Verification / reproduction status
The user would then review and edit the generated engineering note before publication.
A possible workflow:
Select conversation fragment → Generate structured note → Privacy/IP check → User review/edit → Explicit contribution agreement → Publish
The contributor would explicitly confirm that they have the right to contribute the selected material and accept the applicable contribution license.
This must be clearly separated from ordinary ChatGPT privacy settings and from permission to use conversations for model training. It is an intentional act of publication by the user.
Negative results should be first-class knowledge
This is one of the most important aspects of the proposal.
Engineering knowledge suffers from survivorship bias.
Successful designs and experiments are much more likely to be documented than unsuccessful ones. Yet a properly documented negative result can be at least as useful as a successful result.
Knowing that a particular approach fails under specified conditions — and knowing why — may prevent hundreds or thousands of other engineers from independently repeating the same dead end.
Therefore, a failed experiment should not be discarded or hidden.
The knowledge base could explicitly support statuses such as:
Proposed → Calculated → Simulated → Experimentally verified → Independently reproduced → Conditionally valid → Failed → Superseded → Obsolete
A failed solution would remain searchable together with its experimental conditions and identified failure mechanism.
A small real-world example
A recent engineering conversation in ChatGPT illustrates the type of knowledge that could be preserved.
The subject was an old circuit concept for switching a two-channel audio system between stereo and mono by connecting the virtual summing nodes of two operational-amplifier channels.
A physical prototype had originally been built decades ago using an electromechanical reed relay. It worked successfully despite ordinary 5% resistors.
The circuit was recently reconstructed and investigated in simulation.
The investigation proceeded as follows:
- The original circuit was reconstructed in SPICE.
- Component mismatch was deliberately introduced.
- The circuit was tested with resistor mismatch as large as 10%.
- A finite switch resistance of 100 Ω was introduced.
- The ideal switch was then replaced with a real J176 JFET model.
- The simulated JFET had an ON resistance of approximately 325 Ω, yet the fundamental stereo/mono summing principle continued to work.
- With a fast gate transition, however, a substantial switching transient appeared.
- The gate transition was then deliberately slowed to approximately 100 µs.
- The narrow switching transient was strongly reduced.
- This experiment isolated capacitive gate feedthrough / charge injection as the relevant mechanism rather than a failure of the summing topology itself.
- An electromechanical relay was therefore reconsidered.
- A small hermetic REK11 relay was subsequently identified whose datasheet explicitly specifies operation at very small signal currents and uses gold-plated contacts.
- For this particular application, the old electromechanical solution therefore appears technically preferable to the semiconductor switch.
The valuable engineering knowledge is not merely the final recommendation to use a relay.
The useful knowledge object is the entire chain:
hypothesis → simulation → deliberate degradation → observed failure → diagnostic experiment → identified mechanism → alternative solution → datasheet verification
If only the successful final solution were recorded, the negative J176 experiment — arguably the most informative part of the investigation — would disappear.
The next engineer might repeat exactly the same experiment.
Evidence should be explicit
Every engineering claim should carry information about the evidence supporting it.
For example:
Hypothesis → Analytical calculation → Simulation → Single physical experiment → Reproduced physical experiment → Independent reproduction
Simulation and measurement should never silently become equivalent.
Likewise, a plausible explanation generated by AI should remain visibly different from an experimentally established mechanism.
This would make the system fundamentally different from a conventional discussion forum.
A reader should be able to determine immediately whether a statement is:
- someone’s hypothesis;
- an analytical result;
- a SPICE simulation;
- a bench measurement;
- a manufacturer’s specification;
- a reproduced result;
- or a result independently reproduced by several contributors.
The role of AI
AI could make such a knowledge base economically practical.
The engineering community would provide the knowledge and perform reproduction, criticism, and peer review.
AI could perform much of the otherwise expensive editorial and maintenance work:
- converting conversations into structured engineering notes;
- extracting experimental conditions;
- identifying missing information;
- detecting duplicate or closely related contributions;
- linking related experiments;
- identifying contradictory results;
- separating hypotheses from measurements and simulations;
- tracking component lifecycle and obsolete parts;
- detecting newer datasheet revisions;
- generating searchable metadata;
- identifying potentially sensitive or proprietary material before publication;
- translating contributions while preserving technical terminology;
- suggesting experiments capable of reproducing or falsifying a claim.
Importantly, AI should not decide that an engineering claim is true.
Its role should be to organize evidence, preserve provenance, identify inconsistencies, and make the accumulated knowledge searchable.
Technical validity should remain evidence-based and open to reproduction.
Versioning instead of deletion
Engineering knowledge changes.
A component becomes obsolete.
A manufacturer revises a datasheet.
An independent experiment discovers a previously unknown limitation.
A newer component becomes a better solution.
A result initially believed to be universal turns out to apply only within a particular temperature, voltage, frequency, manufacturing, or operating range.
The original contribution should therefore not simply disappear.
Instead, knowledge should evolve through linked versions:
Original result → Reproduction → Discovered limitation → Corrected interpretation → Improved solution → Modern replacement
This preserves not only the answer but also the engineering reasoning that produced it.
In many cases, understanding why an earlier solution was superseded is itself valuable engineering knowledge.
Community verification
Other engineers should be able to reproduce a contribution and attach their results.
A reproduction could contain different:
- component batches;
- manufacturers;
- temperatures;
- supply voltages;
- simulation models;
- PCB layouts;
- measurement equipment;
- operating conditions.
This would gradually transform an isolated observation into characterized engineering knowledge.
Disagreement would also be valuable.
Instead of voting an answer up or down, contributors could provide contradictory evidence with documented conditions.
The system could then expose the boundary:
Works under conditions A–C; fails under conditions D–F.
That is considerably more useful to an engineer than a popularity score.
Privacy and intellectual property
The contribution mechanism should be deliberately conservative.
A conversation may contain information that the user owns alongside information that they do not have the right to publish.
The AI could therefore perform a pre-publication review looking for:
- personal information;
- employer-confidential information;
- credentials or identifiers;
- third-party copyrighted material;
- proprietary documents;
- material whose ownership is unclear.
The user would see the proposed public document before anything leaves the private conversation.
Only the explicitly selected and approved material would be published.
Why ChatGPT is unusually well positioned to do this
Storage is not the difficult part of creating an engineering knowledge base.
The difficult part is converting thousands or millions of differently written laboratory notes, discussions, simulations, failures, corrections, and observations into information that another engineer can actually discover and evaluate.
Historically, that requires enormous human editorial effort.
LLMs fundamentally change the economics of this problem.
Engineers could continue working naturally: discussing a problem, uploading measurements, testing hypotheses, running simulations, discovering mistakes, and correcting conclusions.
Only when something useful emerges would the engineer choose to contribute it.
The AI could then reconstruct that sometimes messy investigation into a concise, structured, searchable engineering record while preserving the underlying evidence and provenance.
The community provides the engineering knowledge.
AI provides much of the indexing, normalization, cross-referencing, and maintenance.
The infrastructure itself is comparatively inexpensive.
It should eventually extend beyond electronics
Engineering would be a natural starting domain because experiments, specifications, boundary conditions, and reproducibility are already central to engineering practice.
But the underlying architecture is not specific to electronics.
The same concept could eventually apply to:
- mechanical engineering;
- materials science;
- chemistry;
- manufacturing;
- software engineering;
- instrumentation;
- optics;
- acoustics;
- laboratory methods;
- repair and diagnostics.
The common object is not a particular discipline.
It is reproducible practical knowledge.
The larger idea
This should not become another forum.
It should not become another Wikipedia.
It should not attempt to replace textbooks, datasheets, scientific literature, or existing engineering communities.
It could instead preserve a category of knowledge that those systems handle poorly:
small, practical, evidence-backed engineering discoveries and failures.
In that sense, the best description may be:
A Collective Engineering Laboratory Notebook
A searchable, versioned repository not merely of facts and final answers, but of:
problems → hypotheses → experiments → evidence → failures → explanations → boundary conditions → reproducible solutions
Engineering advances not only by discovering what works.
It also advances by remembering, with sufficient precision, what did not work and why.