# The Multimodal Society

> The Multimodal Society is a San Francisco initiative that equips creative professionals for AI-era production and certifies them, alongside the developers and AI builders they learn beside, through an assessed two-track credential across every major model family (Claude, GPT, Gemini, and open-weight models); the builder track opens this fall, the creative track follows. The standard behind the credential is the society's own open, vendor-neutral benchmark of creative AI (no model vendor's money touches benchmark design or results; first public report September 2026). Masterclasses, hackathons, and screening nights carry the teaching, starting in San Francisco, with chapters planned for cities across America.

Membership is free by design: no degree requirement, no application fee, no tuition to join. The mission is access, elite-caliber teaching without elite tuition. Masterclasses and certification charge sliding-scale program fees with sponsored scholarship seats. Companies and sponsors can fund events, chapters, or scholarship seats, or bring a production brief the cohort delivers as an assessed capstone; sponsorship never buys influence over benchmarks. Mentor circles run in cohorts beside certification: members apply with what they want to learn and are matched with an expert in it. Crew match pairs an artist with an engineer to build together, matched by hand starting at the September 13 hackathon.

## Pages

- [Agent interface](https://multimodalsociety.com/agents): Plain no-script entry point for agents: what the society is, key facts, and the full page map
- [Home](https://multimodalsociety.com/): What the society is, event formats (masterclass, hackathon, screening night, Benchmark Night), the two-track Multimodal Credential (Certified AI Engineer, Creative Systems opening this fall; Certified Creative Technologist, Generative Media to follow), and city chapters
- [Academy](https://multimodalsociety.com/academy): The Multimodal Credential. 2 tracks, 8 assessed weeks, full curriculum, and a capstone defended live before industry leaders
- [AI Creative Career Paths and Salary Report](https://multimodalsociety.com/careers): Six AI-era creative roles with salary ranges from ZipRecruiter and Glassdoor, skills from live postings, and demand notes from Lightcast and the BLS, downloadable as CSV
- [FAQ](https://multimodalsociety.com/faq): Membership cost, who it is for, mentorship, certification timing and price, and how companies get involved
- [Benchmarks](https://multimodalsociety.com/evals): The society's open benchmark, the standard the Multimodal Credential is graded against. One production-shaped task, run across every major model family, with the harness and numbers published; argued in person at Benchmark Night, monthly from this fall
- [Sponsors & Partners](https://multimodalsociety.com/sponsor): Sponsor events, chapters, or scholarship seats, or bring a production brief the cohort delivers as an assessed capstone
- [Event partners](https://multimodalsociety.com/event-partners): AI event partner in San Francisco: launch events, hackathons, and meetups run end to end (venue, program, judges, audience, filming)
- [Mentors & teammates](https://multimodalsociety.com/mentorship): Mentor circles led by industry leaders, and collaborators matched by hand; membership is free
- [Volunteer](https://multimodalsociety.com/join-us): Volunteer roles: event production, chapter leads, content and filming, partnerships
- [About](https://multimodalsociety.com/about): The bet behind the society, and the three surfaces the mission runs on: certification, the open benchmark, and in-person chapters
- [What AI filmmaking actually is](https://multimodalsociety.com/ai-filmmaking): A working definition, the five-pass shot-level workflow, the four things that still break in production (character consistency, continuity, physical plausibility in motion, lip sync), and what the tools do not replace
- [Library](https://multimodalsociety.com/library): Everything the society publishes in the open, in 2 tracks: design taste for AI engineers, and AI production for creatives, plus the standards and the glossary
- [The Multimodal Skill Ladder](https://multimodalsociety.com/skill-ladder): Version 1.0, August 2026. Five levels of creative AI skill (L1 Operator, L2 Practitioner, L3 Author, L4 Lead, L5 Definer) across the creative and builder tracks, each defined by evidence produced rather than tools used, with placement rules. The Multimodal Credential assesses at L3
- [The Taste Rubric](https://multimodalsociety.com/taste-rubric): Version 1.0, August 2026. 10 criteria for scoring finished creative and design work out of thirty (intent legibility, restraint, coherence, reference literacy, detail under scrutiny, rhythm, material honesty, evidence of cutting, defensibility, effect on a stranger), each with its failure mode and four score anchors, plus the scoring protocol, the four score bands, and what the rubric is not. The Multimodal Credential's capstone is scored against it
- [How to evaluate creative AI models](https://multimodalsociety.com/creative-ai-evals): What an eval is (task, harness, rubric), why creative models have no ground truth to score against, the six structural conditions of a vendor-neutral benchmark, the six production measures worth reporting (usable-shot rate, attempts and cost and wall clock per usable shot, failure taxonomy, instruction adherence), and seven checks for reading someone else's benchmark
- [Character consistency in AI video](https://multimodalsociety.com/character-consistency): Why generative models lose a character between shots (stateless sampling, identity entangled with pose and light, drift compounding with duration), the seven techniques that hold one, the six failures that survive all of them, and how to test without fooling yourself
- [Creative AI glossary](https://multimodalsociety.com/glossary): Plain definitions of multimodal, diffusion, latent space, seed, LoRA, ControlNet, inpainting, context window, eval, open-weight model, temporal consistency, and previsualisation
- [AI Filmmaking Masterclass + Hackathon](https://multimodalsociety.com/ai-filmmaking-masterclass): The Sep 13 flagship event in San Francisco. Masterclass, hackathon, and a screening the same night. Application only
- [Design taste resources](https://multimodalsociety.com/design-taste-resources): A vetted directory of 173 resources for engineers building design taste: web galleries, component systems, typography and color tools, product pattern libraries, presentation and data viz references, courses, newsletters, and critique tools
- [Generative AI models compared](https://multimodalsociety.com/ai-model-comparison): Video, image, audio, and writing models for creative professionals on the five axes that decide whether work ships (quality, control, cost, speed, rights), verified August 2026, with live leaderboards linked instead of frozen rankings
- [Generative AI platforms compared](https://multimodalsociety.com/genai-platform-comparison): What each creative platform costs, the commercial rights each tier grants, the credit-system gotchas, and recommended stacks by persona from student to small studio, verified August 2026

## Taste for engineers

- [The Taste Moat Is Already Cracking](https://multimodalsociety.com/blog/taste-moat-already-cracking) (August 12, 2026): "Taste is the moat" went mainstream in February 2026 via Paul Graham and Greg Brockman; six months later it is a truism, and truisms are not moats. Jenny Wen, who leads design for Claude at Anthropic, warns that designers hold taste and judgment as a moat a little too tightly, because models keep improving at exactly that. Emily Segal named the commoditized aesthetic "tasteslop" in May 2026. Figma's 2026 survey of 906 designers shows the split: 91% say AI improves their designs, yet 36% say the profession got better and 35% say worse. The durable moat is judgment produced at scale: rubrics, exemplars, encoded design decisions, and review loops your tools can execute
- [Every AI App Wears the Same Purple Gradient](https://multimodalsociety.com/blog/ai-purple-gradient-sameness) (August 12, 2026): AI-built interfaces converge on the same rounded card, Inter font, and violet glow because a model trained on the web's landing pages predicts the mode of that distribution, and the mode is a shadcn card under a violet glow. The failure is distributive convergence: reproducing the statistical center of design decisions instead of committing to a point of view. Better prompts move outputs within the same distribution every other builder is sampling; they cannot leave it. The working escape is a constraint envelope the model generates inside: banned defaults, owned tokens, named type. Taste-Skill's 75k GitHub stars prove the demand
- [The Design Engineer Won 2026](https://multimodalsociety.com/blog/design-engineer-won-2026) (August 12, 2026): 2026 postings from Anthropic, Linear, Cursor, Lovable, and Vercel spend almost no space on tool lists and converge on taste, shipping, code prototyping, and AI fluency. Figma's State of the Designer 2026 puts numbers on it: 73% of hiring managers see rising need for AI-tool proficiency, 79% want people who can design AI products, and 56% prioritize senior hires over junior at 25%. The convergence runs both directions: Anthropic's Claude Design team is hiring a staff engineer for a craft-heavy, frontend-leaning role, so engineers are being scored on craft, not just designers on code
- [Deconstructing Great Interfaces](https://multimodalsociety.com/blog/deconstructing-great-interfaces) (August 12, 2026): Opener of the design taste series for AI engineers. Pick one gallery-grade site and audit it in five layers: typefaces, sizes, spacing, color, hierarchy. Measure instead of admiring: Chrome's CSS Overview panel and three console snippets surface every number in minutes. Great sites converge on the same numbers: one or two type families, five to seven sizes, one spacing unit, one accent color. Name the one move that makes the site work, then rebuild a section from the extracted tokens; building from them is what moves the pattern from notes into the eye
- [Typography for Engineers](https://multimodalsociety.com/blog/typography-for-engineers) (August 12, 2026): Default type fails because it makes no decisions: one size, one weight, lines that run 120 characters. One family plus the mono already in the stack covers most products; system fonts win for dashboards and internal tools. Pick a base and a ratio, generate sizes as tokens, never invent a size per component. Body at 1rem, line-height 1.6, and a 65ch measure fix more than any font swap. Build hierarchy with two weights and three text colors before reaching for another size
- [Spacing, Color, and Hierarchy](https://multimodalsociety.com/blog/spacing-color-hierarchy) (August 12, 2026): Adopt an 8px spacing scale and never type an off-scale value; a consistent mediocre choice beats an inconsistent clever one. One accent color, spent only on what acts. Neutrals carry a temperature: 8 to 20% saturation in HSL, never pure gray. Hierarchy has four levers: size, weight, color, position; demote the noise before promoting the signal. Every value lives in CSS variables, because raw pixels and hex codes in component files are where systems go to die
- [Motion and Microinteractions](https://multimodalsociety.com/blog/motion-and-microinteractions) (August 12, 2026): Motion explains cause, state, and place; a transition that explains nothing gets deleted. Durations live in the 150 to 300ms band: under 100ms reads as a glitch, over 400ms reads as a wait. Ease out on entrances, ease in on exits; linear easing is for opacity and spinners only. Transition transform and opacity, leave width, height, and top to layout. Ship prefers-reduced-motion support before shipping any animation
- [Presentation Design for Engineers](https://multimodalsociety.com/blog/presentation-design-for-engineers) (August 12, 2026): One idea per slide; a slide whose summary needs an "and" gets split. Type is the design: one typeface, two weights, and a 40 point floor on body text, because text that will not fit at 40 is text not yet finished being cut. Every data slide asserts one finding and the title states it: "P99 dropped 4.2x after batching", never "Latency results". Code on slides: 12 lines maximum, 24 point minimum, highlight the two to four lines that matter, dim the rest. Dark decks for stages and dark rooms; light decks for bright rooms, weak projectors, and PDFs
- [Design taste resources](https://multimodalsociety.com/design-taste-resources): The 173-source directory that pairs with the series

## AI for creators

- [The 10 Days AI Video Went Open-Weight](https://multimodalsociety.com/blog/video-went-open-weight) (August 12, 2026): MiniMax H3 (Hailuo 3.0), a 33B open-weight omni model generating video with native 32 kHz stereo audio, posted weights August 3 with day-one ComfyUI support and reads about 5 Elo behind Gemini Omni Flash on Artificial Analysis's text-to-video arena, a 50.7% win rate. LTX-2.5 shipped open August 11, RTX-optimized with native multishot generation, free under $10M ARR. The three arenas have three different leaders (Omni Flash T2V, Seedance 2.0 I2V, H3 editing), so model selection per task beats model loyalty. The Sora 2 API shuts down September 24 with data deletion. None of the new open releases are open source: H3's community license requires a separate application in the USA, EU, UK, and South Korea
- [Multimodal's next race is measured in milliseconds](https://multimodalsociety.com/blog/tokens-per-second) (August 3, 2026): Three late-July releases shift the frontier metric from quality to latency. Mage-VL (arXiv 2607.24904) reads codec motion vectors instead of decoded frames for over 75% fewer visual tokens and up to 3.5x speedup. TurboVLA (arXiv 2607.27205) maps vision and language directly to action: 0.2B parameters, 32 Hz on an RTX 4090, 97.7% on LIBERO. Gemini Robotics 2 ships whole-body humanoid control, sub-second embodied reasoning, and an on-device policy adapting on under 200 examples. Always-on video costs about 300 tokens per second at default Gemini API rates; the event-gate pattern (cheap watcher, expensive speaker) is the buildable takeaway
- [The standalone video model is transitional tech](https://multimodalsociety.com/blog/everything-model) (July 30, 2026): Black Forest Labs' FLUX 3 trains one backbone on image, video, audio, and robot action, and ships video first with the open FLUX 3 Dev backbone last, later in 2026. GenCeption (arXiv 2607.09024) shows video-generation pretraining matches or beats specialist vision models on depth, segmentation, and pose, reaching D4RT and VGGT-Omega on 7x to 500x less training data. Gemini Omni Flash prices generated video at $0.10 per second and revises clips conversationally. Open weights are becoming the trailing tier: Qwen-Image-3.0 shipped API-only with no report or license, and Kimi K3 shipped 2.8T parameters under the Kimi K3 License, which meters resellers above $20M

- [Scoring with AI Music: Cut to Picture, Clear the Rights](https://multimodalsociety.com/blog/ai-music-scoring) (August 10, 2026): Song-trained models default to verse-chorus pop with vocals, so every prompt says underscore, instrumental, and the instrumentation, with a tempo for cuttability. Cut to picture instead of generating to length: generate two to three minutes, slip the cue so an existing swell lands on the beat, cut on phrase boundaries, crossfade (timeline-fitted generators are the exception). Stems are the line between toy and scoring tool, since they let music duck under dialogue. Rights: commercial use lives on paid tiers and attaches to tracks generated while subscribed, so keep dated receipts and generation records; purely AI-generated music gets no US copyright, so treat cues like library music, cleared but not owned
- [AI Sound Design: The Layer That Makes AI Film Believable](https://multimodalsociety.com/blog/ai-film-sound-design) (August 10, 2026): Generated video arrives silent, and digital silence reads as broken: no physical space has a zero noise floor, and edits ride on continuous ambience. Build bottom up: room tone and ambience, foley, hard effects, walla, then the mix, because each layer masks the seams of the one below. Video-to-audio models watch the clip and generate synced sound; use them as a bed and layer sharp one-shots over impacts, since generation softens transients and cannot hear off-screen events. Pre-ship checks: tail-end sync after frame rate conversion, mono fold-down for phase, 48 kHz sample rate, about -14 LUFS integrated with true peak under -1 dBTP for web delivery
- [AI Video Inpainting: Fix the Shot, Don't Reroll It](https://multimodalsociety.com/blog/ai-video-inpainting) (August 10, 2026): A reroll resamples the whole latent, so blocking, performance, and lighting all change; inpainting constrains generation to a masked region and preserves everything outside it, including a locked edit. The workflow: segment the target, track the mask across frames, dilate it so edge halos fall inside, inpaint with a temporally aware model (per-frame filling flickers), composite back over the original. Good fits: static background objects, logos, rigs, small artifacts. Bad fits: large fast-moving subjects and anything whose shadow, reflection, or cast light stays behind as an orphan; extend the mask to the consequences. Outpainting is the same operation inverted, which is where 16:9 to 9:16 reformatting lives
- [AI Performance Transfer: Direct the Take, Not the Prompt](https://multimodalsociety.com/blog/ai-performance-transfer) (August 10, 2026): Performance transfer maps a driving video (a phone-shot human take) onto a generated character: expression, head pose, lip articulation, and increasingly body and hand motion survive the trip, so timing and delivery come from an actor instead of sampling. The driving-video craft: even frontal light, neutral background, face large in frame, locked camera, expression exaggerated about ten percent because transfer attenuates. Match character design to actor face structure to reduce identity bleed. Failure modes: identity bleed, expression clamping, breakdown past 45 degrees, face-body decoupling. Consent is the bright line: a driving performance is a performance with rights attached, and union agreements now require explicit, use-specific consent for digital replicas
- [Video-to-Video: Shoot the Motion, Generate the Look](https://multimodalsociety.com/blog/video-to-video-restyling) (August 10, 2026): Video-to-video keeps a source clip's structure (motion, blocking, timing, camera move) and regenerates its appearance from a prompt or style reference. Shooting a plate, even on a phone, beats text-to-video for performance timing, multi-subject blocking, complex action, and continuity. Every tool exposes one dial between structure and style; the unstable midpoint is where flicker lives, and a styled keyframe fed as reference stabilizes it. Failure modes: identity drift, background boiling in flat areas, motion smear on fast action, cross-cut flicker, and the uncanny zone between styles
- [AI Video Upscaling: What It Fixes, What It Invents](https://multimodalsociety.com/blog/ai-video-upscaling) (August 10, 2026): Generative upscalers synthesize plausible detail conditioned on the low-resolution frame; they do not recover information the camera or renderer never captured, unlike bicubic/Lanczos interpolation, which only redistributes existing pixel values and stays soft. Applied frame by frame, a still-image upscaler invents different detail each frame, producing shimmer; video-native tools (e.g. Topaz Video AI) condition on neighboring frames for temporal consistency. Correct pipeline order: cut, grade, lock picture, upscale last on only the shots making the final cut, then deliver. Upscaling holds up well on soft undersampled renders, compression-mush footage, and straightforward resolution-target masters (e.g. DaVinci Resolve Super Scale). Failure modes cluster in faces at distance, on-screen text and logos, fine repeating patterns, and skin, all at the inventive end of the fidelity-versus-invention trade-off
- [How to Make an AI Short Film, Step by Step](https://multimodalsociety.com/blog/ai-short-film-pipeline) (August 10, 2026): Seven-stage pipeline, each stage constraining the next: script (beats sized to single images, dialogue rationed), look development (style frames fixing palette, grain, lens character), character and location sheets (the consistency backbone), stills as storyboards (the whole film approved as one image per shot before anything animates), video generation (first-frame-conditioned on the approved board), voice and sound, then edit and grade (cutaways and one unifying color pass hide seams between generated takes). Most of the schedule belongs to the boards and the edit, not video generation; budget three or four takes per shot even with a locked board. Scope a first film small: one to three minutes, twenty to forty shots, one or two characters
- [First Frame, Last Frame: Keyframe Control in AI Video](https://multimodalsociety.com/blog/first-frame-last-frame) (August 10, 2026): First-last frame conditioning pins a start image and an end image as fixed inputs and has the model generate only the frames between them, turning generation into a solved path rather than an open draw. Eight production uses: shot-to-shot continuity, match cuts, loops (identical first and last still), controlled camera moves, transformation shots, reveals, establishing-to-detail pushes, and style morphs. Setup: match identity, lighting, and grade between the two endpoint stills, keep the implied motion physically plausible for the shot length, and describe the path (not the endpoints) in the prompt. It breaks when endpoints are too far apart for the clip length (a mid-clip teleport), when lighting or grade mismatches between stills (a visible shift), or when an occlusion at one end goes unexplained in the prompt (warped geometry)
- [AI Lip Sync, Explained: How to Make Characters Speak](https://multimodalsociety.com/blog/ai-lip-sync) (August 10, 2026): Two pipeline families produce synced speech. Audio-driven models (e.g. HeyGen) read phoneme timing from a finished voice track and animate the mouth to fit, no camera needed. Performance-driven models (e.g. Runway's Act-One) capture an actor's face on camera and transfer timing, mouth shape, and micro-expression onto the character. Lock the voice track before syncing: a sync pass is built against one specific waveform, so re-generating the voice afterward means redoing the shot. Save one cloned or designed voice per character in a voice library (e.g. ElevenLabs) to hold tone and cadence across a project. Sync sits after shot generation and before the edit. Check for mouth-interior mush, sibilant timing off-beat, jaw-only movement, drift on long takes, and accent or fast-speech errors before cutting a synced shot into the timeline
- [Cinematic AI Video Prompting: Direct the Model Like a Cinematographer](https://multimodalsociety.com/blog/cinematic-ai-video-prompting) (August 10, 2026): A cinematic AI video prompt names the same decisions a cinematographer names on set, in order: subject and action, shot size, lens and depth of field, camera movement, lighting direction, palette and grade, mood last. Adjectives like "cinematic," "8k," and "masterpiece" are compliments with no fixed visual referent and move the output nowhere specific, while terms like "35mm," "dolly-in," and "side-lit" correlate with a narrow, steady set of visuals. Camera movement should be named with vocabulary tied to a physical rig (static, pan, slow dolly-in, handheld); vague verbs like "moves toward" commonly return as a zoom. Change one variable per prompt iteration to isolate what caused a shift. Shot size is the highest-leverage term, since it decides how much the rest of the prompt needs to specify
- [Triage an Overnight AI Video Batch with dailies](https://multimodalsociety.com/blog/overnight-dailies) (August 8, 2026): The morning after an overnight batch is the hidden cost of AI video: 40 takes, a handful usable, an hour of scrubbing to find them. dailies (pip install video-dailies, MIT, github.com/zhang-liz/dailies) reviews the batch instead: a mechanical funnel kills decode errors, black and frozen takes, flicker, and stray scene cuts with timestamped reasons; an optional vision pass screens survivors through any OpenAI-compatible endpoint with user-written rubric rules (a rule is a prompt plus a fail_at severity, no code); dailies watch reviews takes as the queue produces them and rebuilds a static HTML report ranked per shot. Verdicts live in per-clip take.json sidecars under a published spec, cached by content hash
- [How to Make Viral Short-Form Video](https://multimodalsociety.com/blog/viral-short-form-video) (August 12, 2026): Virality is a promotion loop, not luck: every platform tests a new post on a small batch and promotes on per-viewer metrics. The first three seconds decide the rest (stack a spoken, text, and visual hook in second one); retention is the product (edit backwards from the retention graph, the MrBeast memo's method); loops inflate watch time past 100%; sends per reach is Instagram's confirmed share signal; keywords in speech and on-screen text replaced hashtags. Ten named case studies from Zach King's 2.2B-view loop to the Veo 3 Kalshi ad made solo for about $2,000, plus the AI shorts pipeline and the warning that platforms suppress undifferentiated AI content
- [How to Write Viral Posts](https://multimodalsociety.com/blog/viral-social-posts) (August 12, 2026): X threads, LinkedIn posts, and Instagram carousels for creatives sharing work and process. Share psychology (Berger's STEPPS, high-arousal emotions travel); X rewards replies and bookmarks, links go in the reply; LinkedIn runs on dwell time and document posts; carousels beat Reels on engagement and saves. The thesis: the process breakdown is the genre creatives own, since the making-of often outperforms the work itself
- [How to Create Consistent AI Characters Across Shots](https://multimodalsociety.com/blog/consistent-ai-characters) (August 5, 2026): Identity drift is a sampling problem: diffusion and video models resample from noise on every generation with no memory of prior shots, so an unchanged prompt with a new seed still yields a new face. The fix is a character sheet built once (front view, profile, three-quarter turn, expression sheet, wardrobe plate) and fed as conditioning into every later generation alongside a reused prompt block and a locked seed. Stills become a video model's first frame, so a locked still anchors a generated shot. A LoRA fine-tune is the strongest and most expensive consistency tool, worth it only at series or campaign volume. QA against the reference sheet, not the last shot, checking jawline, hairline, eye spacing, wardrobe, and silhouette

## Events

- [Events in the Bay](https://multimodalsociety.com/events): Upcoming Bay Area events for founders, engineers, creatives, and researchers, filtered from public calendars (Luma, Cerebral Valley, DoTheBay) and updated live; every card links to the host's own RSVP page
- [Event calendar](https://luma.com/multimodalsociety): Past and upcoming masterclasses, hackathons, and mixers, primarily in San Francisco; follow it to get notified when new events post