OpenAI · Launch Video Breakdown: Hook, Pacing & Motion Design

Scene-by-scene timeline & spoken transcript

1. 0:00 – 0:04

   The Hook

   Introducing GPT-Rosalind

   “(No spoken dialogue — A consistent, driving rhythm with pizzicato synths and sub-bass begins.)”

   On screen

   Accelerate scientific research with GPT-Rosalind

   Camera

   Static, centered text on a white background.

   Motion

   Simple fade-in for the title text.

2. 0:04 – 0:40

   Product Reveal

   Interactive Scientific Query

   “This is a life sciences model. It's built to support research across biology, drug discovery, and translational medicine. It can help you analyze data, synthesize evidence, and prioritize targets. Let's say you're working on an asthma program and you want to compare IL-33, TSLP, and IL-1RL1 for asthma target prioritization. You can use the local package in inputs as supporting evidence. The model will then process this information and provide a detailed analysis.”

   On screen

   What will you build? asthma program Ask Codex anything, @ to add files, / for commands, $ for skills Custom GPT-Rosalind Extra High BLAST a sample protein sequence and return the top hits using the NCBI BLAST skill Assess LRRK2 as a Parkinson's target via parallel evidence lanes with subagents Optimize a sample 96-well luminescence inhibition assay with CMPD-327 Connect your favorite apps to Codex Compare IL33, TSLP, and IL1RL1 for asthma internal_assay_results.csv Asthma Research Open with Numbers assay_id assay_name assay_family sample_group target endpoint unit n mean_value cl_low cl_high p_value prefer_flag notes A1 IL2 cytokine suppression eosinophilic severe asthma donors IL33 IL33_IL1RL1_suppression_pct percent 18 69.0 65.0 73.0 2e-05 True pass Strong suppression in type-2-high donor IL2 culture. A1 Primary IL2 cytokine suppression eosinophilic severe asthma donors IL33 IL33_IL1RL1_suppression_pct percent 18 61.0 56.0 61.0 7e-05 True pass Strong suppression with better cross-donor consistency. A2 Bronchial biopsy type-2 signature reduction efficacy type-2 high severe asthma biopsies TSLP TSLP_CCL26_signature_reduction_pct percent 16 55.0 50.0 60.0 4e-05 True pass Mechanistically clean CCL2 effect, but less durable activity. A2 Bronchial biopsy type-2 signature reduction efficacy type-2 high severe asthma biopsies TSLP TSLP_CCL26_signature_reduction_pct percent 16 52.0 47.0 57.0 3e-05 True pass Best composite signature shift across the biopsy frag. A3 Epithelial-IL2 co-culture viral mimic efficacy polyIC-stimulated airway epithelium plus autologous ILC2 IL33 IL33_IL1RL1_reduction_pct percent 12 66.0 61.0 71.0 3e-05 True pass Biospy response is present but narrower than the lung. A3 Epithelial-IL2 co-culture viral mimic efficacy polyIC-stimulated airway epithelium plus autologous ILC2 IL33 IL33_IL1RL1_reduction_pct percent 12 64.0 59.0 69.0 3e-05 True pass Captured upstream alarmin-driven amplification. A4 Whole-blood innate challenge perturbation safety healthy donor whole blood IL33 host_defense_perturbation_index index 20 0.21 0.17 0.24 0.005 False pass Lower is better; modest perturbation under asthma model. A4 Whole-blood innate challenge perturbation safety healthy donor whole blood TSLP host_defense_perturbation_index index 20 0.17 0.14 0.2 0.003 False pass Lowest whole-blood perturbation among the ligands. A5 Biomarker responder rate translational FeNO-high moderate-to-severe asthma TSLP responder_rate fraction 32 0.67 0.57 0.76 0.004 True pass Responder defined as >30% complete biomarker r. A5 Biomarker responder rate translational FeNO-high moderate-to-severe asthma TSLP responder_rate fraction 32 0.72 0.63 0.81 0.003 True pass Best FeNO-high responder rate with strong carryover. A6 Biomarker responder rate translational FeNO-low moderate-to-severe asthma IL33 responder_rate fraction 41 0.41 0.3 0.53 0.008 True pass Lower response suggests type-2 enrichment of the s. A6 Biomarker responder rate translational FeNO-low moderate-to-severe asthma IL1RL1 responder_rate fraction 41 0.39 0.28 0.5 0.011 True pass Lower FeNO-low carryover is consistent with a narrow. Compare IL33, TSLP, and IL1RL1 for asthma target prioritization. Use the local package in "inputs" as supporting evidence. Worked for 2m 32s Using only the local inputs bundle, the rank is TSLP > IL33 > IL1RL1 for broad asthma target prioritization. TSLP is the best lead because it has the strongest overall execution package, not just good biology. In the tractability table it has the best biomarker observability, safety headroom, and competitive validation scores, with Clinical ready assay readiness; in the internal assays it also has the best FeNO-high and FeNO-low responder rates ( 0.72 and 0.53 ) and the lowest host-defense perturbation ( 0.17 ). That lines up best with the TPP's emphasis on breadth, safety, and simple biomarker-guided development. IL33 is the best second choice if the portfolio wants a sharper epithelial-stress or viral-exacerbation angle. It shows the strongest upstream efficacy in the ILC2 and viral-mimic assays ( 69% and 66% ), and its biomarker strategy explicitly fits epithelial injury/alarmin-stress biology. The problem is execution: the package flags IL-33 as low-abundance and context-dependent, and the low-FeNO carryover is weaker ( 0.41 ) than TSLP, so it looks more segment-specific than broad. IL1RL1 ranks third from this package alone. The receptor-axis story is mechanistically coherent, but the package is consistently narrower: lower efficacy across the three efficacy assays ( 63/59/61 ), the highest innate-perturbation index ( 0.24 ), the weakest low-FeNO carryover ( 0.39 ), and a more complex sST2/receptor-occupancy biomarker setup. Unless external genetics or differentiation strongly favors ST2 biology, it is the least attractive development path here. Confidence is moderate, not high. The metadata explicitly says this is a synthetic internal evidence bundle and that these assays should support, not replace, human genetics or orthogonal target-disease evidence.

   Camera

   Simulated screen recording of a user interacting with the GPT-Rosalind interface. The camera pans and zooms slightly to highlight different sections of the UI and generated text.

   Motion

   UI animation, text typing simulation, smooth camera pan and zoom, data visualization (bar charts) appearing dynamically.

3. 0:40 – 1:20

   Feature Teaser

   Deep Dive with Sub-Agents

   “Now let's dig deeper. Let's use the Life Science Research plugin to gather public evidence from genetics, cohort follow-up, regulatory context, target disease evidence, clinical precedent, and disease-relevant literature. We'll spawn sub-agents for each lane of evidence to avoid bias. The research router workflow splits this into six independent evidence lanes and launches separate explorers. Each sub-agent focuses on a specific area, like human genetics, and uses relevant skills to gather information. This process ensures a comprehensive and unbiased analysis.”

   On screen

   Ask for follow-up changes Let's dig deeper. Let's use the Life Science: Research to gather public evidence from genetics, cohort follow-up, regulatory context, target disease-evidence, clinical precedent, and disease-relevant literature. Spawn subagents for each lane of evidence to avoid bias. Working for 2m 9s I'm using the research-router workflow to split this into six independent evidence lanes and launch separate explorers Thinking Compare asthma targets Pascal... asthma program Compare IL33, TSLP, and IL1RL1 for asthma target prioritization using PUBLIC evidence only. Important: Do not read or use the local /inputs package; independence matters for this lane. Use the Life Science: Research plugin skills relevant to your lane. Read only the skill docs you need. Return a concise report with: 1. What you checked 2. Key findings by target 3. Main caveats and conflicts 4. Lane-level ranking or takeaway 5. Source URLs or canonical identifiers for the main public sources Lane: human genetics. Focus on direct human genetic support for asthma and closely related allergic/respiratory phenotypes. Use skills such as opentargets-skill, gwas-catalog-skill, gtex-eqtl-skill, ensembl-skill, and gnomad-graphql-skill if useful. Report key variants/loci, mapped genes, effect directions when available, and whether ... Show more 10:24 PM Working for 8s I'm gathering the public genetics evidence for IL33, TSLP, and IL1RL1 and will stay strictly outside /inputs . First I'm reading the relevant Life Science: Research skill docs, then I'll query public sources for asthma and closely related Working for 2m 3s I'm using the research-router workflow to split this into six independent evidence lanes and launch separate explorers for each one. They'll work from public sources only so the local inputs bundle doesn't anchor their conclusions prematurely. I'll reconcile their outputs against the local evidence afterward. Spawned 6 agents Created Pascal (explorer) with the instructions: [mention]life-science-research[/plugin][life-science-research@openai-curated] Compare IL33, TSLP, and IL1RL1 for a... Created Lovelace (explorer) with the instructions: [mention]life-science-research[/plugin][life-science-research@openai-curated] Compare IL33, TSLP, and IL1RL1 for a... Created Peirce (explorer) with the instructions: [mention]life-science-research[/plugin][life-science-research@openai-curated] Compare IL33, TSLP, and IL1RL1 for a... Created Lovelace (explorer) with the instructions: [mention]life-science-research[/plugin][life-science-research@openai-curated] Compare IL33, TSLP, and IL1RL1 for a... Created Beauvoir (explorer) with the instructions: [mention]life-science-research[/plugin][life-science-research@openai-curated] Compare IL33, TSLP, and IL1RL1 for a... Created Gauss (explorer) with the instructions: [mention]life-science-research[/plugin][life-science-research@openai-curated] Compare IL33, TSLP, and IL1RL1 for a... The lane agents are running now: genetics, cohort follow-up, regulatory, target disease-evidence, clinical precedent, and literature. I'm keeping the public-evidence work isolated while I prepare the synthesis frame for the local package. I'm waiting on the first lane results to land; once a few complete I'll start reconciling them rather than waiting for all six at once. No lane has finished yet. I'm keeping the waits coarse to avoid polling noise and will surface the first completed evidence lane as soon as it returns. Thinking 1 set of 3 tasks completed 1. Define evidence lanes and launch subagents 2. Collect and synthesize public evidence from each lane 3. Integrate with local package and produce final prioritization 6 background agents Ask for follow-up changes or to log an agent Custom GPT-Rosalind Extra High Work locally

   Camera

   Continued simulated screen recording, focusing on the typing of new queries, the display of agent spawning, and the progress of tasks.

   Motion

   UI animation, text typing simulation, dynamic display of task progress and agent activity, subtle background color shift.

4. 1:20 – 2:34

   Product Reveal

   Synthesized Prioritization & Visual Artifacts

   “Using six independent public evidence lanes from the Life Science Research plugin, then reconciling them with the local package, the overall rank stays TSLP > IL33 > IL1RL1. TSLP remains the lead because it is the only pathway here with clear asthma approval and the most de-risked public clinical package. IL33 is the strongest second choice if you want the best human-genetics signal and the clearest epithelial-injury/viral-exacerbation biology. IL1RL1 is third overall. The only real disagreement across lanes was the IL33 vs IL1RL1 ordering. Pure genetics favored IL33; cohort follow-up and a narrow read of asthma clinical precedent favored IL1RL1. The integrated decision still favors IL33. Great! Let's make some visual artifacts like: rank/evidence heatmap, GWAS sentinel and association-count plot, internal assay panel.”

   On screen

   partial report now, rather than letting six long-running searches keep accumulating latency. Messaged 5 agents The first completed lane is in Open Targets-based disease evidence puts IL33 > TSLP > IL1RL1 on asthma association strength, but still favors TSLP since asthma-specific clinical de-risking is weighted. I'm pulling the remaining lanes now so we can see whether genetics, cohorts, and literature agree or diverge. Three lanes are now in and they're aligned: TSLP is strongest once clinical and regulatory de-risking are counted, IL33 keeps surfacing as the mechanistic and genetics-heavy alternative, and IL1RL1 trails on execution despite a credible biology story. I'm waiting on the genetics, cohort, and clinical-precedent lanes to test whether that pattern holds up when we surface human genetics and replication. Clinical-precedent came back slightly differently than the literature lane: it still favors TSLP clearly, but it places IL33 ahead of IL1RL1 on public additional prevalent follow-up. The subagent's phase 2 signal is cleaner than the literature's phase 2 signal. I'm waiting for the dedicated genetics and cohort replication lanes to decide whether that clinical tie-break should matter for final prioritization. The genetics lane breaks from the earlier TSLP-first pattern: on pure human genetics it ranks IL33 > IL1RL1 > TSLP mostly because IL33 has the clearest protective loss-of-function story and IL1RL1 has a protective sST2 signal, while TSLP's locus-to-direction link is weaker in the public datasets. I'm waiting on the cohort-replication lane to see whether real-world cohort follow-up strengthens or dilutes that genetics signal. All six public evidence lanes are in. I'm shuttling the subagents down and merging the lane-level disagreements into a single recommendation, with the local package used as the final tie-break where public evidence conflicts. Closed 5 agents Closed Lovelace (explorer) Closed Gauss (explorer) Closed Beauvoir (explorer) Thinking 2 out of 3 tasks completed 1. Define evidence lanes and launch subagents 2. Collect and synthesize public evidence from each lane 3. Integrate with local package and produce final prioritization Ask for follow-up changes Worked for 9m 25s Using six independent public-evidence lanes from the Life Science: Research plugin, then reconciling them with the local package, the overall rank stays TSLP > IL33 > IL1RL1. TSLP remains the lead because it is the only pathway here with clear asthma approval and the most de-risked public clinical package: tezepelumab has a broad severe-asthma label, strong phase 2/3 precedent, and the literature supports breadth beyond classic eosinophilic disease, including reassuring viral-challenge biology. The local bundle still breaks the tie in its favor because it has the best biomarker observability and safety/translation package in tractability_and_safety.csv (line 3) and the strongest FeNO-high/low responder performance in internal_assay_results.csv (line 15) and internal_assay_results.csv (line 18). Public anchors: Open Targets asthma, Tezspire label, NAVIGATOR, SOURCE. IL33 is the strongest second choice if you want the best human-genetics signal and the clearest epithelial-injury/viral- exacerbation biology. It won the genetics lane because of the protective splice-loss signal at rs146597587, and it was essentially tied with TSLP on target-disease evidence. It loses to TSLP because asthma clinical precedent is still phase- 2-level and the biomarker package is less straightforward, which the local bundle also flags in biomarker_strategy.csv (line 2) and tractability_and_safety.csv (line 2). Public anchors: rs146597587, Itepekimab asthma phase 2, IL-33 rhinovirus exacerbation paper. IL1RL1 is third overall. It has real strengths: the cleanest cohort follow-up in this run, a protective-expression signal at the locus, and a credible phase 2 asthma signal with interest in lower-eosinophil disease. But the locus is partly shared with IL18R1, the receptor/soluble ST2 biology is harder to interpret and monitor, Open Targets asthma evidence is lower than IL33/TSLP, and the local package penalizes assay complexity and narrower breadth in biomarker_strategy.csv (line 4) and tractability_and_safety.csv (line 4). Public anchors: rs72823641, FinnGen variant page, Astegolimab phase 2. The only real disagreement across lanes was the IL33 vs IL1RL1 ordering. Pure genetics favored IL33 ; cohort follow-up and a narrow read of asthma clinical precedent favored IL1RL1. The integrated decision still favors IL33 for #2 because the asthma TPP rewards broader disease evidence, clearer mechanism, and clearer differentiation in epithelial-stress/exacerbation biology, while IL1RL1 still carries more ambiguity and a messier biomarker story. If you want, I can turn this into a one-page scoring table with per-lane weights and an explicit scorecard. If you want the practical call: lead TSLP ; keep IL33 as the upstream, genetics-backed backup; keep IL1RL1 as the clinically interesting but execution-heavier watchlist target. If you want, I can turn the six-lane evidence into a one-page scorecard with explicit lane weights. Great! Let's make some visual artifacts like: - rank/evidence heatmap - GWAS sentinel and association-count plot - internal assay panel Thinking Ask for follow-up changes Full access GPT-Rosalind Extra High Work locally Internal Assay Panel A1 ILC2 cytokine suppression eosinophilic severe asthma donors A2 Bronchial biopsy signature Type-2 high severe asthma biopsies A3 Viral-mimic co-culture polyIC-stimulated airway epithelium plus autologous ILC2 A4 Innate challenge perturbation healthy donor whole blood A5 FeNO-high responder rate FeNO-high moderate-to-severe asthma A6 FeNO-low responder rate FeNO-low moderate-to-severe asthma TSLP IL33 IL1RL1 percent index

   Camera

   Continued simulated screen recording, showing the final synthesized report and then transitioning to dynamically generated bar charts.

   Motion

   UI animation, text typing simulation, dynamic display of final report, smooth transition to data visualizations (bar charts) with animated bar fills.