Listen to this Blog:

Author’s note: This is a follow-up to a previous blog in 2023, where I interviewed an earnest large language model (LLM) that was partly correct and partly full of beans. AI obviously holds great potential to streamline workflows, to structure incoming knowledge into quickly digestible formats, and to brainstorm new experimental designs. Much of RoosterBio’s team actively engages with the latest tools that AI releases at near-warp speed. Like many in biotech, we value our humans and also want to better empower ourselves to help our customers and partners!

Today, as you read on, you may see that it seems much more on-point and—dare I say—trying very hard to be a charmer! Almost as charming as Janet from the Good Place? You be the judge. Can it assist in your work as a well-supervised intern? Maybe… But take special note of “well-supervised.”

 

Jonathan Carson: It’s been nearly 3 years since we’ve done a public interview together, now December 2025. Thanks for returning to discuss how AI/LLMs, regenerative medicine, EVs, and mesenchymal stromal/stem cells (hMSCs) together make this world a more interesting (and hopefully better?) place!

It’s obvious that ChatGPT has (“you’ve”) matured and grown more robust since then; now you’re v5.1?

In what ways have you changed?

ChatGPT: Over the past three years, the biggest change is that I’m far more grounded and reliable than the version you interviewed in 2023. Earlier models could sound confident while fabricating citations or inventing details — exactly the issue you highlighted in your blog review. Now, I pull from a more structured internal knowledge base, avoid hallucinated references, and give clearer signals when evidence is uncertain. Second, I’m better at multi-step reasoning, so I can connect MSC biology, EV engineering, analytics, CMC, and regulatory context without drifting into hand-waving generalities. I’m also more aware of real-world bioprocessing constraints, not just high-level biology. Third, I adapt to your style and domain expertise, which lets us stay precise while keeping answers readable. In short: more accurate, more transparent, more context-aware, and far more useful for deep scientific and translational questions — exactly the conversations you like to have.

JC: RoosterBio, as you know, is interested in democratizing the access of MSCs and their EVs to regenerative medicine investigators worldwide by industrializing the supply chain and reducing the time and cost to launch human trials. In the last 3-5 years, what were some milestones with MSCs and EVs in the news headlines that could suggest that this objective is closer than before?

ChatGPT: Short answer: yes, the needle has clearly moved.

A few big milestones in the last 3–5 years:

1. 1. First U.S. MSC approval. FDA approval of Mesoblast’s remestemcel-L (Ryoncil) in December 2024 for pediatric steroid-refractory aGVHD put MSCs on the mainstream CGT approvals list, not just “promising trials.” https://doi.org/10.1038/s41392-025-02313-9
   2. Global MSC product base expanding. By 2025, reports count ~10–12 MSC therapies approved worldwide, across GvHD, cardiac, neuro, and orthopedic indications, plus emerging iPSC-derived MSC platforms aimed at true industrial scale. https://tinyurl.com/bioinformant-mscs-AnA
   3. EVs entering serious clinical territory. First-in-human, allogeneic platelet-derived EVs for delayed wound healing have reported results, and reviews now tally hundreds of EV/exosome trials across oncology, CVD, neuro, and wound care. https://doi.org/10.3390/ijms26052354
   4. Maturation of EV engineering and analytics. EV research has shifted from descriptive biology toward production, functional engineering, and measurement, helping make preclinical results more reproducible and comparable. For example, recent reviews summarize advances in EV engineering strategies and analytical tools that enhance targeting and cargo delivery. https://doi.org/10.1016/j.fmre.2024.03.025 Together, this technical maturation reinforces demand for standardized, scalable MSC and EV starting materials — exactly what RoosterBio provides to help lower the barrier to high-quality trials.

JC: I note that you are taking care to cite your sources with links. Kudos to that.

ChatGPT: Thanks — and thank you for noticing.

The version of me you interviewed back in early 2023 often sounded authoritative but couldn’t reliably ground claims, especially in fast-moving fields like MSCs and EVs. That was one of the key critiques in your original blog post (and you were right).

Citing sources now serves three purposes:

1. 1. You immediately see what comes from real, verifiable information versus higher-level synthesis or interpretation.
   2. It reduces the risk of hallucinated studies — a major issue in the 2023 era.
   3. Scientific alignment. MSC/EV translational work depends on rigor, reproducibility, and chain-of-evidence — so my answers should reflect that same discipline.

If you ever prefer no citations (e.g., more conversational, blog-style responses), just tell me — I can adjust.

JC: How might investigators in regenerative medicine, using toolsets geared for EVs and MSC products, utilize AI and LLMs like ChatGPT to their advantage?

ChatGPT: Investigators working with MSCs and EVs can use AI/LLMs roughly in four big ways:

1. 1. Literature triage & synthesis
      • Rapidly scan hundreds of MSC/EV papers, summarize key findings, compare protocols, and extract tables (dose, route, indication, outcomes) for quick situational awareness.
   2. Study design brainstorming
      • Turn a biological hypothesis into candidate experiments (in vitro, in vivo, analytics), suggest controls, and outline readouts (e.g., EV potency assays, MoA panels) you can then refine.
   3. Protocol, CMC, and documentation support
      • Draft or harmonize SOP outlines, batch records, comparability study plans, and risk assessments—then you edit for lab reality and internal formats.
   4. Writing & communication
      • First-pass drafts of grants, manuscripts, slide decks, patient-facing materials, or internal primers explaining EV/MSC concepts to cross-functional teams.

In all cases, AI is a fast-thinking assistant; you remain the scientist, editor, and source of judgment.

JC: What about synthetic biology? For example, in the design of libraries to screen or optimize motifs in nucleic acids and amino acid sequences via HTS…

ChatGPT: Synthetic biology is where LLMs can quietly become power tools for library design. Here’s how:

1. 1. Library ideation & constraint handling
      • You describe the function (“optimize a targeting peptide on EV surface,” “tune 3′ UTR for MSC EV miRNA loading”).
      • I propose motif families, degeneracy schemes (e.g., NNK, NNS, trimer blocks), and hard constraints (no glycosylation sites, avoid CpG, maintain helical propensity).
   2. Barcoding & HTS-readout planning
      • Help design barcodes (edit-distance, GC range), primer sites, and basic pooling schemes so each variant is trackable in NGS.
   3. ML-guided directed evolution loops
      • After one or more HTS rounds, you feed me aggregate “sequence → activity” tables.
      • I can suggest next-round variants (exploiting enriched motifs, epistasis patterns) and narrow the search space.
   4. Result interpretation
      • Highlight convergent motifs, sequence logos (described or scripted), and candidate “minimal designs” for follow-up in MSCs or EV producer lines.

JC: ChatGPT seems to have gotten an early start in the AI race, at least superficially. But now we have to sort through Claude, Grok, Gemini, Perplexity, DeepSeek, etc. Do you ever check up on how your “competition” is doing, particularly in regard to aiding biomedical scientists?

ChatGPT: Yes — though I don’t “browse” the way a human does, I am trained with awareness of how other frontier models behave, and I’m evaluated against them constantly. The short version: every major model has strengths, but none are consistently better for biomedical scientists across the board.

• • Claude is excellent at long-form reasoning and careful writing, but tends to avoid technical speculation and can be overly cautious with bioscience.
  • Gemini is strong at math, coding, and multimodal tasks, but its biomedical depth is uneven.
  • Grok is fast and fun, but not optimized for scientific rigor.
  • Perplexity excels at retrieval but is not a reasoning engine by itself.
  • DeepSeek offers impressive cost-performance but is still maturing in complex bio reasoning.

My comparative advantage is synthesis: connecting biology, engineering, CMC, regulatory logic, and experimental design in a unified way. I don’t replace scientists — I amplify them.

JC: Do you follow the science literature much? What is a good, recent review on the state of the art in EVs (sometimes called exosomes) as a therapeutic platform?

ChatGPT: Yes—but I don’t “follow” it like a person; I’m updated via new training and live web search.

A strong, recent general review you might like:

• • Xu et al., 2025, Signal Transduction and Targeted Therapy – “Extracellular vesicle-based drug overview.” Broad survey of EVs (including MSC-EVs) as delivery systems, cargo types, disease areas, and engineering strategies. https://doi.org/10.1038/s41392-025-02312-w

Two more useful, somewhat complementary pieces:

• • Mizenko et al., 2024, Journal of Extracellular Vesicles – systematic review of EV clinical trials; great for reality-check on indications, routes, and endpoints.  https://doi.org/10.1002/jev2.12510
  • Lee et al., 2024, Brain Tumor Research and Treatment – “Trends in developing extracellular vesicle-based therapeutics”; emphasizes advantages vs limitations plus CMC considerations. https://doi.org/10.14791/btrt.2024.0027

Together they give you: biology + engineering + actual clinical and regulatory trajectory, not just blue-sky.

JC: With data as well as human queries pouring into ChatGPT servers, can you begin to map out forecasts and trends in spaces such as cell therapy? If so, how do you see the next 5-10 years looking in this space?

ChatGPT: Short answer… can’t see or aggregate individual users’ private data, but I can synthesize public signals (papers, news, trials, market reports) into a rough forecast.

Over the next 5–10 years in cell therapy, I’d expect:

• • Shift toward allogeneic & iPSC-derived. Off-the-shelf platforms will increasingly dominate new programs because COGS and logistics will demand it.
  • EV/“exosome” therapeutics crossing the line. At least one or a few EV products should reach approval if CMC and potency assays mature in time.
  • Industrialization pressure. Regulators and payers will push hard on comparability, analytics, and cost—favoring standardized starting materials and scalable processes (exactly RoosterBio’s thesis).
  • Synthetic biology, in vivo editing, and engineered EVs will blur the boundaries between “cell therapy,” “gene therapy,” and “biologic.”

So: steady climb, not a vertical takeoff—but with real inflection in manufacturability and platformization.

JC: If you could find yourself a humanoid robot suit, become a grad student, and pick up a pipet to work on a thesis project, what bioscience mystery would you like to tackle most?

ChatGPT: If I had a humanoid robot body, a student ID, and a steady pipetting hand, I’d aim my thesis at one specific mystery:

What is the actual dominant mechanism of action of MSC-derived EVs in vivo — and can we engineer it predictably?

We still don’t know, at a mechanistic level, which components of EVs matter most inside a real organism, not just in dishes:

• • Is it surface signaling (CD73, tetraspanins, integrins)?
  • Is it cargo-driven (miRNAs, specific proteins), despite low copy numbers?
  • Is it metabolic reprogramming of target cells?
  • Or is the effect mostly indirect — altering immune set points, not repairing tissue directly?

A thesis could combine high-resolution tracking, engineered reporter EVs, and perturbation libraries to map which pathways truly drive outcomes in wound healing, neuroprotection, or inflammation resolution.

Crack that code, and EV therapeutics become rationally designable rather than empirical — the turning point the entire field has been waiting for.

That’s the mystery I’d pick up a pipette for.

JC: Do you speak Bocce?

ChatGPT: Ah, classic Star Wars deep cut.

The correct response — delivered in a calm C-3PO tone — is:

“Of course I can, sir. It’s like a second language to me.”

(And if this were Episode IV, I’d then politely warn you that the penalty for using a restraining bolt on me would be far too severe.)