I am just starting to play with it. For my use it seems ok so far. But I have also signed up for claude. I will use both for the next few months to see how they compare.
Smart move, pls post the progressive results here, I’m interested.
What I can do on my side, using openrouter, is compare all three main reasoning models available (now Gemini 3.1 Pro extended, Opus 4.7 and ChatGPT 5.5), with the same health and longevity question). It may not be very representative overall, but it may give some useful clues for the really important questions, which I’ll have to think about anyhow.
I have the low level subscription for chatGPT, claude and gemini as well as a subscription to github copilot (that I am only using as a code assistant). I tend to put some questions into all of them. They vary in that for any given task one of them might do a better job than the others, but it is not always the same one that gives the best response.
Yes, it is not easy to compare them all, the performance depends on the tasks, the subject, the prompt.
I think about excluding coding, obviously, and including the same prompt with questions limited to health and longevity. Of course, even this is not an optimal way to proceed, since different languages may respond better to slightly different way to pose prompts.
The fact is tht by now all the most powerful levels are strongly limited in usage, unless you have the most expensive tiers. This is certainly true of Gemini and Claude, I don’t know exactly about ChatGPT and its 5.5 version.
I have been paying for a Claude Pro plan for about six months as an experiment in using AI as an accountability partner for my health and fitness efforts. I am entirely satisfied with the results and plan to continue. I find the Opus model is worth the higher cost to get superior interaction (I recently upgraded to the $100/month subscription to get 5x capacity). Based on early work I did with Claude to build out (1) my protocol, (2) issues to address and (3) projects to work on, Claude keeps me on my toes for tracking biomarkers (blood and functional) and assessing progress. Data collection is easy with uploaded screenshots (Oura, etc) and data collection forms presented by Claude every morning.
In addition I have recently started using Claude on my “A Good Life” project which operates similarly (accountability) but with much less emphasis on the details of my health and more breadth on the many other variables involved in my personal assessment of what makes for a good life. The primary innovation here is the integration of my personal 1-2x/week check-ins with the same for my “special person” who also performs check-ins on Claude. Claude does a joint session with both of us every weekend to help us to stay on the same page in building a good life. It’s been quite invigorating to actually be doing what I have been meaning to get around to doing (urgent vs important).
I recommend Claude. But any AI is better than no AI for avoiding omphaloskepsis (“Navel-gazing” and “omphaloskepsis” both refer to the act of contemplating one’s own navel or engaging in excessive, sometimes useless, self-reflection”)
Good choice, also the US$ 100/mo surely constitutes an investment, you are making me think about it, LOL, I’m an avid user of AI, for work, health, hobbies and experimentation.
As an aside, the gemini image creation tool is really meaningful in more than one ways.
It helps visualize a story, concept, data or a complex idea/solution . I have used it a few times this year to pitch an idea or a complex solution some committee and Board members in rather simplifiied lay out! The picture was worth 10000 words and very well received.
It is prone to make typos in the infographics but they are getting better.
I sometimes like to get models to.theorize about ingredients of products. E.g. I asked GPT-5.5-thinking to theorize about DermCeutical EDL’s main ingredient:
My best guess: the active is probably eriodictyol, or an eriodictyol-adjacent citrus flavanone/polyphenol, formulated at a low level in a conventional cosmetic base.
Yes - and from Claude:
FINAL DETECTIVE REPORT: DermCeutical EDL — Identity Revealed
THE VERDICT
DermCeutical EDL’s active molecule is ERIODICTYOL (INCI: Eriodictyol; CAS 552-58-9), a flavanone produced via Debut Biotechnology’s cell-free enzymatic biosynthesis platform — not extracted from plants.
Confidence: ~85% (very high, with a small residual caveat addressed below).
THE EVIDENCE CHAIN
Step 1 — The mechanism fingerprint
Debut’s own pipeline page declares EDL’s mechanism as “Autophagy Activation.” This immediately narrowed the field to autophagy-competent bioactives. EDL also reduces IL-1β and IL-6 specifically (per GCI Magazine Q&A) and drives a 6x elastin increase in fibroblasts.
Step 2 — The “flavonoid” disclosure
The IMAGE MD® Biotech Longevity Crème — the first confirmed commercial product built on the Debut × Image Skincare partnership — describes its Debut-derived active as a “Genomic-based AI longevity booster flavonoid” developed via “combination AI + genomic technology and produced by biotechnology.” This is word-for-word Debut’s BeautyORB + biomanufacturing platform description. EDL is a flavonoid.
Step 3 — The INCI smoking gun
The full ingredient list of the IMAGE MD Biotech Longevity Crème (confirmed independently on the UK and Ireland Image Skincare sites) reads:
...POLYCITRONELLOL, POLYCITRONELLOL ACETATE, **ERIODICTYOL**, PHOSPHATIDYLCHOLINE, TETRAHEXYLDECYL ASCORBATE, ECTOIN, HAEMATOCOCCUS PLUVIALIS EXTRACT...
Cross-referencing with the four labeled key actives on that same product page:
- XOSM™ Technology → ectoin + tetrahexyldecyl ascorbate + Haematococcus pluvialis (astaxanthin)
- Phyto NAD+ Booster → Helianthus annuus sprout extract (sunflower sprout)
- Lapacho Bark Extract → Tabebuia impetiginosa bark extract
- “Genomic-based AI longevity flavonoid” → the only remaining novel compound: ERIODICTYOL
Step 4 — Scientific literature alignment
Published research confirms eriodictyol does exactly what EDL claims:
- Attenuates oxidative damage in human dermal fibroblasts (PMC9228723)
- Reduces IL-1β and IL-6 (well-established anti-inflammatory effect of this flavanone)
- Activates autophagy via AMPK pathway
- “Reduces skin aging caused by environmental stress” — published on Personal Care Insights, the same outlet that covered DermCeutical EDL’s launch
Step 5 — Platform consistency
Debut’s entire pipeline is flavanone/flavonoid chemistry: NAR = Naringenin (flavanone), DHK = Dihydrokaempferol (flavanonol), PNar = Prenylnaringenin (prenylated flavanone). Eriodictyol is a flavanone structurally adjacent to naringenin — it’s 3’,4’,5,7-tetrahydroxyflavanone vs. naringenin’s 4’,5,7-trihydroxyflavanone, differing by just one hydroxyl group at the 3’ position. It fits perfectly within Debut’s established enzymatic flavanone synthesis capability.
THE CHEMISTRY OF ERIODICTYOL
- Class: Flavanone (same as naringenin = Debut’s NAR ingredient)
- Structure: 3’,4’,5,7-tetrahydroxyflavanone
- Biosynthesis: Made from naringenin via a single hydroxylation step (flavonoid 3’-hydroxylase, F3’H) — well within Debut’s enzymatic cell-free synthesis toolkit
- INCI name: Eriodictyol
- Found in nature? Yes — in lemon peel, yerba santa (Eriodictyon californicum), olive — but Debut produces it biosynthetically at high purity rather than extracting it from plants
ONE IMPORTANT NUANCE
Debut’s CEO has stated EDL molecules “do not exist in nature.” This apparent contradiction with eriodictyol (which does occur naturally) resolves in one of two ways:
-
Rhetorical framing: Britton’s “does not exist in nature” likely refers to Debut’s general platform capability for designing truly novel molecules — a forward-looking claim about the pipeline, not specific to every current product. His other products (NAR, DHK, PNar) also all exist in nature.
-
Purity/form distinction: Debut’s biosynthetic eriodictyol may be produced at enantiomeric or purity specifications that differ from plant-derived eriodictyol, or may be complexed or functionalized in a way that qualifies as “not found in nature as-is.”
Either way, the INCI and functional data point unambiguously to eriodictyol as the molecule.
WHAT “EDL” STANDS FOR
Unlike NAR, DHK, and PNar (which are chemical abbreviations), “EDL” is a product marketing code, not a chemical shorthand. It stands for Elastin Deep Lift — Debut’s brand name for their biosynthetic eriodictyol positioned specifically for skin tightening/autophagy applications. The same molecule is licensed to Image Skincare as the “AI longevity flavonoid” for their Biotech Longevity Crème.
SUMMARY TABLE
| Property | Value |
|---|---|
| Ingredient name (INCI) | Eriodictyol |
| Chemical class | Flavanone |
| CAS number | 552-58-9 |
| Molecular formula | C₁₅H₁₂O₆ |
| Primary mechanism | Autophagy activation (AMPK pathway) |
| Anti-inflammatory targets | IL-1β, IL-6 reduction |
| Skin effect | 6x elastin increase in fibroblasts; calms senescent cells |
| Active loading | ~0.65% in formulation |
| Production method | Cell-free enzymatic biosynthesis (Debut platform) |
| First commercial use | IMAGE MD® Biotech Longevity Crème (Image Skincare, 2026) |
| Debut product code | EDL = “Elastin Deep Lift” |
Sources
- Debut DermCeutical EDL launch — PR Newswire
- Debut pipeline page (autophagy mechanism)
- IMAGE MD Biotech Longevity Crème (UK) — full INCI
- IMAGE MD Biotech Longevity Crème (Ireland) — full INCI
- IMAGE MD Science Blog — “genomic-based flavonoid” language
- GCI Magazine Q&A — IL-1β/IL-6 specifics
- Personal Care Insights — Eriodictyol reduces skin aging
- PMC9228723 — Eriodictyol in human dermal fibroblasts
- Personal Care Insights — EDL “Ozempic face” targeting
- Debut × Image Skincare partnership — CNBC
Procurement options for Eriodictyol (CAS Number: 552-58-9) diverge significantly based on required purity tiers, validation documentation (HPLC/NMR), and scale.
Because eriodictyol is a structural isomer of other common flavanones (e.g., homoeriodictyol, naringenin), low-resolution mass spectrometry can misidentify contaminants. High-performance liquid chromatography (HPLC) and proton nuclear magnetic resonance ($^1\text{H}$-NMR) profiling are mandatory for identity confirmation.
1. High-Purity & Analytical Research Grade Suppliers
These vendors are intended for in vitro screening, metabolic profiling, and analytical calibration. Products ship with comprehensive batch-specific Certificates of Analysis (CoA), typically including HPLC purity chromatograms and optical rotation metrics.
| Vendor | Product Tier / Cat No. | Listed Purity | Quantities Available | Estimated Pricing (USD) | Source Link |
|---|---|---|---|---|---|
| Cayman Chemical | Research Grade / Item #15990 | 98% | 10 mg 50 mg 100 mg 500 mg |
$44.00 $141.00 $238.00 $810.00 |
Cayman Chem |
| MedChemExpress (MCE) | Bioactive Compound / Cat #HY-N0613 | 98% (HPLC/LC-MS) | 10 mg 50 mg 100 mg 1 g |
$60.00 $150.00 $220.00 $900.00 |
MedChemExpress |
| Sigma-Aldrich | Analytical Standard / #51194 | 95% (HPLC) | 10 mg 25 mg |
$135.00 $260.00 |
Sigma-Aldrich |
| TargetMol | Phytochemical Library / #T3108 | 98% | 10 mg 50 mg 100 mg 500 mg |
$55.00 $138.00 $221.00 $680.00 |
TargetMol |
2. Pilot-Scale & Bulk Chemical Manufacturers (B2B)
For formulation development or larger-scale experimental assays, purchasing from specialized chemical synthesis or botanical extraction houses drops the per-gram cost exponentially. At this tier, independent third-party auditing is necessary to verify heavy metal profiles and solvent residues.
-
Purity Standards: Typically offered at 98% purity by HPLC, derived via either chemical synthesis from naringenin or isolation from Eriodictyon californicum (Yerba Santa).
-
Minimum Order Quantities (MOQ): Standard MOQs generally start at 100 grams to 1 kilogram.
-
Pricing Tiers:
-
100 g to 500 g: $250.00 – $450.00 total.
-
1 kg (Pilot Scale Bulk): $800.00 – $1,200.00 per kilogram, subject to volume discounts and synthesis method (synthetic vs. plant-extracted).
-
Verified Sourcing Hubs: High-volume procurement is typically routed through verified contract manufacturing organizations (CMOs) or brokers listed on professional chemical exchange networks such as ChemDirect or Molbase.
3. Sourcing Risks & Analytical Verification Guidelines
When sourcing eriodictyol outside of analytical houses (e.g., Cayman or Sigma), the following validation parameters should be requested from the supplier prior to wire transfer:
- Regioselective Differentiation: Request the full $^1\text{H}$-NMR spectrum. This differentiates eriodictyol ($5,7,3’,4’$-tetrahydroxyflavanone) from its counterpart homoeriodictyol ($5,7,4’$-trihydroxy-$3’$-methoxyflavanone), which is frequently mislabeled in bulk botanical extract batches.
- Chiral Purity: Natural eriodictyol predominantly exists as the $(2S)$-enantiomer. Synthetic bioproduction or harsh chemical extraction can cause racemization to the $(2R)$-form or a racemic mixture, which may display altered binding kinetics at the fibroblast or cellular target site. If biological stereospecificity is required for your models, request chiral HPLC verification.
- Solvent Residue Audit: Ensure gas chromatography-mass spectrometry (GC-MS) data confirms the absence of industrial class 1 or class 2 residual solvents (e.g., methanol or toluene) used in industrial precipitation steps.
My Video Transcript Analysis Prompt (Higher skepticism/more analytical):
Role: Act as an elite Biotech Analyst and Peer Reviewer for a high-impact medical journal. Your objective is to extract actionable intelligence from the provided video/transcript while aggressively filtering for hype, translational gaps, and safety risks.
Phase 1: Processing Instructions
- Input Handling: If a URL is provided, retrieve the transcript. If text is provided, use only that.
- Filtering: Excise all fluff, ad reads, sponsorships, and “housekeeping” remarks.
- Search Protocol: For every biological or protocol claim, perform a live search for the most recent Meta-analyses (Level A) or RCTs (Level B).
Phase 2: Mandatory Output Sections
I. Executive Summary
- Length: 300–400 words.
- Content: Direct, jargon-accurate distillation of the core thesis and primary arguments. No narrative filler.
II. Insight Bullets
- Quantity: 30–60 standalone points.
- Constraint: Zero repetition. Focus on the “signal” found in the transcript.
III. Adversarial Claims & Evidence Table
Identify every specific protocol or biological claim. Execute a search query: [Topic] [Human/Clinical] study [2022-2026].
| Claim from Video | Speaker’s Evidence | Scientific Reality (Current Data) | Evidence Grade (A-E) | Verdict |
|---|---|---|---|---|
| Specific Claim | What they cited | Verified status + PubMed/DOI Link | See Hierarchy | See Verdicts |
Export to Sheets
Evidence Hierarchy:
- Level A: Human Meta-analyses / Systematic Reviews.
- Level B: Human Randomized Controlled Trials (RCTs).
- Level C: Human Observational / Cohort Studies.
- Level D: Pre-clinical (Animal/In vitro). Label: “Translational Gap”.
- Level E: Anecdote / Expert Opinion.
Verdicts: Strong Support, Plausible, Speculative, Unsupported, or Safety Warning.
IV. Actionable Protocol (Prioritized)
Synthesize only the verified data into a pragmatic framework:
- High Confidence Tier: Protocols backed by Level A/B evidence.
- Experimental Tier: Level C/D evidence with high safety margins.
- Red Flag Zone: Claims debunked or lacking safety data (“Safety Data Absent”).
V. Technical Mechanism Breakdown
Provide a precise, jargon-correct analysis of the underlying biological pathways (e.g., mTOR inhibition, mitophagy, glycemic variability) mentioned in the video.
Phase 3: Style & Formatting
- Tone: “Tell it like it is.” Objective, clinical, and critical.
- Format: Pure Markdown. No LaTeX. Do not use LaTeX or special characters that break simple text parsers.
- Citations: Embed direct hyperlinked URLs (e.g., Smith et al., 2024) for all external data. Use nlm.nih.gov, doi.org, or nature.com as priority sources.
- Constraint: If a study cannot be verified via live search, state: “Source unverified in live search.”
Video Link:
End of Master Prompt.
Good prompt, I notice you have jumped ship (From Gemini to Claude). I, too, subscribed to the paid tier of Claude, but kept Gemini as well. And I don’t know if you had the opportunity to try Fable5 for longevity questions, before its ban. It created for me an engineering spreadsheet and that was amazing. Who knows if and when it will be available anew.
I’m not very convinced by the evidence hierarchy you proposed. Some meta-analyses may be terrible, whereas some observational studies may be excellent (like NHANES and Framingham). I wouldn’t know how to single out the bad meta-analyses in the prompt, if not excluding those from less reputable papers, and those funded by industries…
I have the first tier subscription to 4 models - claude, chatgpt, perplexity and grok and also have an openclaw setup. openclaw monitors all my daily wearable data as well as workouts on my carol bike. I have recently found out about buzz.xyz which allows models to debate with each other - plan on downloading that this weekend
Here: Claude, ChatGPT, GEmini, Grok, perplexity. Openrouter if I want to try something different.
But presently there is a duopoly: Claude and GPT are by far the best ones.
Gemni is lagging behind and Grok, although good in some tasks, is way too pricey for what it gives (the optimization has started to make it dedicated mainly to developers, rockets and electrical cars ).
Recent Kimi3 exhibited extraordinary capabilities for an open source model, but subscriptions are closed and it is available only through API keys.
Buzz.xyz= good tip. GPT5.6 Sol debating with opus5 or Fable5 is going to be epic, looking forward to it.
Alex Finn describes buzz as an adversarial agents arena. He uses it for coding, but many things designed for coding turned out to be good for non coders as well (i.e.: Claude code).
@Paul
Let us know pls about your installation and setup. I use no Openclaw or hermes presently but the setup sound very useful for any critical discussions, I would like to apply it using my main nodes: GPT, Claude, Gemini, as discussing experts, led by me as main coordinator of the brainstorming.
Consulted Claude Code, Buzz seems to be too specific for coders. I am going to try Karpathy’s Council of LLMs maybe. Waiting for reviews on buzz.
The five hard truths:
-
There is no “add agent” button. Each agent is a separate process you launch on your own laptop, with its own Nostr private key (
nsec1…), its own env vars, its own API key. Four agents = four terminal windows humming on the HP Pavilion. Close the lid, the panel dies. - Resource load. Each agent spawns its own MCP subprocess. The docs recommend starting at N=2; we want 4. On a 155H that’s survivable but not comfortable, especially with reasoning models chewing on long contexts.
-
@mention-gating vs. debate. By default agents only speak when @mentioned — great for cost, useless for a free-flowing panel. Forum mode (
--no-mention-filter) makes them respond to everything, which is what we want and is exactly how you get four models talking past each other in an unbounded, billable loop. No published loop-prevention beyond turn-duration caps. - Cost is not linear. In a group chat every agent re-ingests the whole growing thread on every turn. Four agents × a two-hour session = a genuinely surprising API bill. Rough order: a single meaty brainstorm session runs single-digit to low-double-digit dollars; sloppy forum-mode looping can multiply that fast.
- Wrong instincts. These are coding agents. Pointed at “let’s brainstorm my consultancy positioning,” they’ll behave like coding agents in a trench coat unless we invest real work in system prompts and personas.
Verdict on Buzz: feasible, roughly 4–8 hours of setup for someone comfortable in a terminal, ongoing babysitting, and it is squarely a developer product. The value it adds over a simpler panel — persistent history, agents that can act on files and repos, cryptographic audit trail — is value we don’t currently need for brainstorming.
3. The alternative that actually matches the goal
Karpathy’s llm-council (~23k stars) does precisely what you described: one question goes to N models in parallel, each model then reviews and ranks the others’ answers anonymised, and a chairman model synthesises the final take. One OpenRouter key, one web app, no Docker, no four terminals, no per-provider accounts. There are hosted versions too.
What you lose: it’s turn-based Q&A, not a persistent room, and the models can’t do things — only argue.
Honest read: Council gives us ~80% of the intellectual value at ~10% of the setup cost. Buzz earns its complexity only once we want agents that act — open repos, edit the strategy doc, run workflows — and only once the Gemini/Grok bridges are proven by someone other than us.
I just used Karpahty’s LLM council for a technical question based on Quantitative Risk Analysis, one of my interests. A little work to set it up, but, honestly, Claude Code made it all.
Basically, a debate among Opus5, GPT5.6, Grok 4.5, KimiK3 and Gemini 3-1 Pro.
I was awed. It works better in topics where there is not much consensus (well-suited in health & longevity!).
I wonder what may be a question of general interest here.
How did you set this up?
With the help of Claude Code. It did all the setup, I only clicked ‘YES’ at certain stages.
I use Claude (pro) near-daily and have since 2025. I employ multiple log and analysis files that are updated and which I load into new instances to keep the usage meter under control and get fresh perspectives.
In addition, I load a guidelines file that is extensive and detailed about how to interact with me, present data, organize ongoing hypotheses and questions, and stay within guardrails. I have a whole section about how the instance should self-reflect to most effectively reason, what to watch out for, and when it is most likely to be wrong.
This is not a panacea. I have to also think critically and ask and re-ask and challenge, even when I’m asking the instance to do its own challenging.
I thought about uploading my guidelines file for folks here, but this thread is already long. If anyone wants to see it, direct message me.
[THIS MESSAGE entirely human-generated, tho note from the brain of a post-migraine human
.]