Gemini 3.1 Flash Live posts 90.8% and 95.9% – Search Live rollout

Gemini 3.1 Flash Live (Google DeepMind): Google is rolling out Gemini 3.1 Flash Live as a real-time audio model focused on more natural back-and-forth plus stronger tool use via function calling, as described in the launch thread.

Video loads on view

It’s now appearing in consumer surfaces—Gemini Live inside the Gemini app and Search Live—and also as a build target for developers inside Google AI Studio, as stated in the rollout note and detailed in the launch post.

• Conversation reliability upgrades: DeepMind frames the practical gains as being better in noisy environments and better at tracking long conversations “so you don’t have to repeat yourself,” according to the launch thread.
• Realtime agents framing: The model is pitched specifically for building voice-and-vision agents with improved quality and latency, using the “step function improvement” claim in the realtime model launch.

Gemini 3.1 Flash Live (Google): Shared benchmark charts position Flash Live as a big step up for speech-to-tools reliability, with ComplexFuncBench (audio) function-calling accuracy at 90.8% in the benchmarks charts.

On the same post, a “Big Bench Audio” chart shows 95.9% speech reasoning for “Gemini 3.1 Flash Live (Thinking High),” placing it near the top of the compared set in the benchmarks charts. The charts also highlight a large gap between “Thinking High” and “Thinking Minimal” (70.5%), which implies an explicit quality/latency trade available to builders, as shown in the benchmarks charts.

Multi‑Shot App (Runway): Runway introduced Multi‑Shot, a web app flow that takes a single prompt and outputs a structured scene with intentional cuts, pacing, cinematic framing, and support for dialogue + sound effects, as described in the Multi‑Shot announcement and surfaced via the web app. It’s being shown through short comedic “scripted moment” prompts—less montage, more scene construction—with examples like awkward character banter and therapy-room setups in the Squirrel and seagull clip and the Monster group therapy clip.

Video loads on view

• Prompt shapes shown today: Dialogue-first awkwardness in the Squirrel and seagull clip and premise-plus-turn structure in the Mice fishing argument suggest Multi‑Shot is optimized for punchline timing (beats, reaction cuts), not only visual spectacle.
• Genre stretch beyond comedy: A darker, cinematic prompt is also showcased in the Toad potion shop scene, implying the same “multi-cut” scaffold can carry mood and tension, not just gags.

AI Selves (Pika): Pika says AI Selves are now in public beta—positioned as an “agentic extension” of a person—available at pika.me and via a new iOS app, per the Public beta announcement. The pitch is less “make a clip” and more “spawn a persistent persona,” which matters for creators building recurring characters and formats.

Video loads on view

Seedance 2.0 (Dreamina/ByteDance): A tight horror beat—an elevator descent with escalating tension anchored by the line “What floor are you going to?”—is being used as a quick genre stress test for pacing, timing, and mood control in very short runtimes, as shown in the Elevator horror clip. It reads like a template for testing whether a model can hold suspense without “doing too much.”

Video loads on view

Kling (Kuaishou): Kling posted a stylized sword-duel clip (“two cat masters face off”) as a motion/impact showcase, emphasizing readable contact moments (sparks on clash) and sustained choreography in a short vertical sequence, as shown in the Sword duel sample. For filmmakers, it’s a quick reference point for how Kling handles continuous action versus cut-heavy edits.

Video loads on view

UNI-1 (Luma Labs): Following up on Directable pitch—multi-reference control—Luma published a set of four “Input 1 + explicit instruction + Output” panels that foreground what “directable” means in practice: keep a character reference stable while changing setting/aesthetic, execute precise material constraints in an architectural render, and even transform an entire landscape into an embroidered patch while preserving composition, as shown in the directable examples.

• Prompt pattern to steal: The screenshots repeatedly use “Use image 1 as character reference…” plus tight art-direction cues (era, palette, lens/film grain) to reduce unintended drift, as demonstrated in the directable examples and described on the Model page.

Nano Banana (image generation): Nano Banana Pro and Nano Banana 2 added 4K image generation, expanding beyond prior 1K/2K choices; the update is framed as a practical lever for better downstream AI video results when using images as keyframes or references, according to the 4K support note.

The UI screenshot shows the new 1K / 2K / 4K selector alongside model choice and batching controls, as visible in the 4K support note, and it also incidentally shows that policy filtering still applies at higher resolutions (a “Generation failed” panel appears in the same capture).

UNI-1 (Luma Labs): Following up on Intelligent pitch—“intent-grounded visuals” positioning—Luma is now showing four concrete samples where UNI-1 keeps dense layouts readable (blueprints, timelines, atlas-style charts) rather than melting into texture, as highlighted in the info design claim and reinforced by the visual set.

The examples matter because they’re the kind of artifacts creatives actually ship—deck diagrams, world-building charts, editorial infographics—where legibility and hierarchy are the whole job; Luma’s claim is that UNI-1 can hold layout hierarchy + typography under heavy information density, per the info design claim and the details on the Model page.

Midjourney (style reference): Another widely copy-pasted code is --sref 3272229711, described as a 2D adult-sitcom animation look in the neighborhood of Bob’s Burgers / The Great North, with some BoJack-like touches, per the style description.

Because it’s tuned for “animation still” readability—flat shapes, clear silhouettes, simple lighting—it’s a practical house style for pitch frames, animatics, and quick scene boards where consistency beats painterly detail, as illustrated in the style description.

Midjourney (style reference): A new shared style reference, --sref 2890513616, is being used to deliberately produce “bad drawing” aesthetics—rough portraits and scribbly fan-art-like silhouettes—framed as “mimic the drawing style of an AI hater,” according to the sref callout.

This is useful when the goal is anti-polish (zines, comedic inserts, storyboard scratch frames, poster marginalia), because it makes “imperfect” a controllable look instead of a failure mode, as shown in the sref callout.

Phota (Phota Labs): Phota’s likeness-focused image model is now described as publicly available, with a pitch that it can generate new photos that “actually look like you (and your pets)” and also edit/enhance real photos (e.g., fix flaws, adjust expression), as laid out in the launch note from Public availability claim and the access pointer in Access link. More detail on where to try it appears via the Studio page, with additional product/onramp framing in Studio and API mention.

• What creators are doing with it: The shared use-cases include short-prompt “photo booth” generations, “style me” with reference images, and “make pro” portrait polishing, as described across Photo booth workflow, Style me reference workflow , and Make pro feature.

No pricing, evals, or a model card are surfaced in the tweets; the evidence here is workflow-level and testimonial (e.g., “made hundreds of photos”).

UNI‑1 Pouty Pals (Luma): A specific “mini‑me” image format is spreading where UNI‑1 generates a 3D chibi version of you sitting in an open palm, with a second hand pressing the cheek into a pout; multiple creators are posting variants and re-sharing the exact prompt, which is doing a lot of the consistency work (scale, hand interaction, expression) as seen in the prompt paste from Exact prompt text and examples from Pouty Pal output, Family Pouty Pal , and Chibi-fied self.

• Prompt payload that matters: The repeating core is “big head small body,” “sitting on an open left palm,” and the “right hand gently pressing its cheek,” which bakes in an interaction constraint that’s harder for many models than a plain portrait, per the shared template in Exact prompt text.

The meme is less about style variety and more about a reliable likeness capture ritual that people can run on any clean selfie.

UNI‑1 Pouty Pal (Luma): A step-by-step “make yours in ~60 seconds” recipe is being shared with explicit claims about what happens to uploaded photos—select UNI‑1, upload a clear front-facing image, paste the prompt, and wait ~30 seconds, per the walkthrough in 60-second how-to and the UI demo in Prompt steps demo.

Video loads on view

• Privacy and rights language: The thread claims uploaded photos aren’t used to train models and that users keep rights to their likeness/output, framed as “Luma confirmed all of this,” according to Privacy claim.

• Mechanical expectation-setting: The same thread positions the output as “hyper-detailed 3D chibi” with “realistic hands” and shallow DoF, matching the recipe described in Output description.

The only hard evidence here is the repeated reproducible prompt + outputs; the privacy assertions are presented as claims in-thread rather than linked to a formal policy artifact.

LTX 2.3 LoRA (LTX): A creator reports training an LTX 2.3 LoRA for a specific character persona (George Costanza) locally on an RTX 5090 in about a day using “AI Toolkit,” framing it as a DIY path to reusable identity/character consistency, according to Home LoRA training note.

There aren’t settings, dataset size, or a reproducible notebook in the tweet, but it’s a clear signal that “character in a box” workflows are moving into consumer-hardware timeframes for some creators.

Phota (Phota Labs): A specific identity workflow is being highlighted where you train separate likenesses for multiple people/pets and then combine them inside Phota’s editor for “family photo” style results, which is positioned as preserving each subject’s details while blending them together, per the walkthrough in Multi-person and pets workflow and the broader capability framing in Model resembles you pitch.

The notable part is the composition step: instead of one prompt trying to juggle multiple identities, the workflow treats each identity as its own asset and then composes downstream.

Clawdbot + Kling + CapCut (workflow): A repeatable affiliate-content loop is being shared as “one consistent persona, endless hooks”—Clawdbot generates the expert persona + scripts, Kling generates the talking-head clips, and CapCut handles captions/retention editing, with affiliate links converting from the bio as described in the workflow breakdown.

• What’s operationally new: The workflow emphasizes constant A/B testing at scale (“same character, different symptom… winners get scaled”) rather than polishing a single flagship asset, per the workflow breakdown.
• Why it matters for creators: It’s an explicit “production line” model—persona consistency + modular hooks—so the creative work shifts toward scripting and thumbnail/hook iteration more than filming, as shown in the workflow breakdown.

Kling + Suno + edit pipeline (workflow): A creator claims a full “high-budget” music video can be produced in under 2 hours for about $200 by combining Kling for video generation (including Motion Control) with Suno for the track, then packaging prompts + a production breakdown for reuse, as outlined in the workflow offer.

Video loads on view

• Proof-of-process angle: The same thread includes a “we flew to Africa… didn’t actually fly” reveal that frames the approach as location/production-value simulation rather than on-set capture, as shown in the green screen reveal.
• Where the full pack lives: The author points people to a Telegram group for the detailed workflow + prompts, per the Telegram group.

Seedance 2 (workflow pattern): A detailed template is being shared that treats video prompting like a shotlist—declare lens/stock/lighting + “face stable” constraints, then specify actions and camera moves in timecoded blocks (0–1s, 1–3s, …) and build multiple sequences that get trimmed in DaVinci Resolve, as described in the base template and expanded in the two-sequence example.

Video loads on view

• Why the structure helps: The example sequences split a scene into “Armor construction” and “Launch,” each with its own cinematic setup and negative constraints (no face morph, no armor floating), per the two-sequence example.

Creator moat discourse: A recurring claim is that product/creative execution is becoming trivially replicable—“run 3 prompts… makes a replica… open-source it”—leaving “taste” (creative direction and judgment) as the defensible edge, as argued in the moat quote.

The thread doesn’t provide the specific repo or implementation details, so treat it as a sentiment signal rather than a verified case study.

Claude Code (workflow): One creator shows an end-to-end loop from “hacking my network” to building a custom home automation app UI, positioning Claude Code as the agentic coding layer that turns device control ideas into working endpoints, per the build note.

Video loads on view

The demo shows a simple phone UI (e.g., Lights/Fan) tied to code functions, suggesting a pattern where creators build bespoke control surfaces for real-world setups instead of relying on off-the-shelf dashboards, as shown in the build note.

Nano Banana (prompt template): A full “citylight poster leaning on brutalist concrete wall” mega-prompt is circulating with unusually strict art-direction rules—brand-palette logic, blob-grid layout system, hero product masked inside a blob, and a hard “text embargo” (only tagline + wordmark), as laid out in the Wall poster mockup prompt and repeated in the Full prompt block.

The point is consistency: it bakes in wall material, dappled tree shadows, lens/DoF, and typography constraints so you can swap [BRAND NAME] and keep the same premium outdoor-OOH look across many renders.

Midjourney (SREF 2885679472): A motion-blur aesthetic is being shared as a reusable “frozen frame from a moving film” look—heavy grain, glowing highlights, smeared motion—framed as “Wong Kar-wai shot through a dirty 35mm lens,” per the Motion blur sref writeup.

If you want the exact prompt scaffolding, it’s packaged in the prompt breakdown page, which also spells out common use cases (album covers, fashion editorials, sports campaigns).

Nano Banana 2 (prompt recipe): A long, production-leaning prompt is being used to turn a 2D character concept into a hyper-real 3D CGI collectible—explicitly calling for physically based materials, Octane/V-Ray-like lighting language, and aggressive error repair (“Fix and reconstruct all hands and fingers… no fused digits”), per the Collectible figurine prompt.

It’s basically a “make it manufacturable” checklist for characters: preserve pose/outfit, remove 2D artifacts, enforce symmetry, and push micro-textures (weave, pores, scratches).

Adobe Firefly (prompt template): A copy-paste macro/product-photo prompt is being shared that generates a miniature world inside a transparent container (decanter/flask/lantern), with explicit lens (85mm), shallow DoF, internal self-lighting, and refraction rules, as written in the Container prompt share.

It’s structured as fill-in slots like [TRANSPARENT CONTAINER] and [COSMIC OR ATMOSPHERIC PHENOMENON], which makes it quick to iterate across themes.

Midjourney (SREF 2873816195): A “neon dream” look is being shared as a repeatable recipe—deep blue shadows, electric edge glow, soft blur and digital noise—positioned for flyers, indie game art, VR worlds, and sci‑fi promos in the Neon dream sref post.

The longer prompt setup is collected in the prompt breakdown page.

Midjourney (SREF 3422279710): A style reference is being pitched as a reliable way to get “expensive” Art Nouveau—hand-drawn texture, graceful linework, muted tones—optimized for covers/editorial/luxury visuals, as described in the Art Nouveau sref post.

The copy-paste base parameters and prompt guidance live in the prompt guide.

Seedance 2.0 (shot-language prompt): A single prompt example is being shared that leans heavily on film grammar—“handheld camera angle,” “realistic lens tremor,” “shallow depth of field,” and slow micro-actions—to keep a calm fantasy close-up stable (child reading while held by a dragon), as shown in the Prompted video example.

Video loads on view

It’s a good example of prompting for pace and micro-motion (page turns, blinking, breathing) instead of big action beats.

Midjourney (SREF 2543866241): A “jet trails tearing through the sky” SREF is being shared with example outputs that read like long-exposure aviation photography—streaked lights, twilight blues, and motion-smeared faces/terrain—per the Jet trails sref post.

It’s a compact code drop: the value is in the reference’s motion language, not a long prompt.

Midjourney (SREF 8006572439): A “cold-tone cyber dark aesthetics” SREF is highlighted as the current most-liked style on a SREF-tracking site, with the post emphasizing teal/blue grading, volumetric fog, wet reflections, and neo-noir tension in the Sref popularity snapshot.

The same post links to the catalog entry on the Sref library page, but treat the ranking as site-specific (not a platform-wide Midjourney leaderboard).

Starlight Precise 2.5 (Topaz Labs): Topaz launched Starlight Precise 2.5 to make GenAI characters/footage more usable—targeting the “plastic” look and common artifacts while upscaling outputs to 4K, with availability called out across Astra plus partner surfaces like HeyGen, Higgsfield, and ComfyUI in the launch post.

Video loads on view

• Where it runs: Topaz points creators to Astra for access, as described on the Astra app page, and frames the rollout as integration-friendly in the launch post.
• Production signal: Topaz also says Precise 2.5 is live for API customers, per the API availability note.

Uncharted Life 2.0 (Topaz Labs community): A shared clip credits a Midjourney + Topaz stack as the finishing workflow—positioning Topaz as the last-mile step to clean up and present AI-generated visuals as a coherent, higher-fidelity sequence, as shown in the creator example.

Video loads on view

The post doesn’t disclose settings (model choice, strength, or any pre-denoise), so it reads as a “this is the polish stage” proof point rather than a replicable recipe.

Topaz Express (Topaz Labs): A simple pattern is being promoted in art threads—share an image, run a Topaz Express upscale, then repost as the “finished” version—explicitly framed as “QT with your Eye art” in the prompt to participate.

This is less about heavy editorial control and more about consistent presentation quality for quick community sharing.

3D Scenes (Freepik): Freepik is pitching 3D Scenes as an “AI 3D photo shoot” workflow—generate a full environment from any image, drop in objects, then move the camera like a real shoot while keeping lighting/detail coherent across angles, as shown in the 3D Scenes demo.

Video loads on view

Developers/teams who’ve avoided full 3D pipelines but still want repeatable coverage (product spins, ad plates, storyboards) now have a more direct “scene-first” interface, with the product entry point available via the 3D Scenes tool.

LagerNVS (paper): A new novel view synthesis method, LagerNVS, is being highlighted as “fully neural real-time NVS,” with the paper page reporting 30+ FPS at 512×512 on a single H100 and a 31.4 PSNR on Re10k, as summarized in the paper page and shared via the demo clip.

Video loads on view

For creative tooling, this points at a near-term path to “move the camera” experiences without a conventional reconstruction step—especially if paired with diffusion decoders for generative extrapolation (noted in the same paper page).

Hybrid photography pipeline: Linus Ekenstam frames the emerging workflow as “simulated inside a 3D environment, enhanced and generated by diffusion,” arguing it keeps the old affordances (camera control) while shedding the painful parts of traditional 3D pipelines, per his future of photography take.

Video loads on view

This lands as a broader pattern around “camera-first” creation—where the scene representation becomes the stable substrate and diffusion becomes the renderer/filler—rather than the usual text-to-image loop.

Wonder 3D (Autodesk Flow Studio): Flow Studio is sharing a prompt-driven loop where each Wonder3D prompt is treated as a “vibe selector” for 3D character direction, then the output is pushed through refinement steps like remeshing and retexturing (including export to other tools), as described in the Wonder3D prompt loop.

Video loads on view

The emphasis is less “one perfect gen” and more “generate directions fast, then clean up where it matters”—which maps closely to game/previz prototyping workflows.

Seedance 2.0 (ByteDance): Seedance 2.0 is now visibly selectable across multiple ByteDance creation surfaces—Dreamina, CapCut’s web Video Studio, and Pippit—via in-product model menus that show different “Fast/+” variants and duration tiers (including <15s vs 15–90s), as documented in the model selector screenshots.

This matters for creators because it signals a single underlying video model being distributed through multiple “studio” front-ends (ads, social edits, short clips, and longer-form assembly), with the practical difference living in the UI constraints (duration, aspect ratio, and reference handling) rather than the model name alone.

Phota Studio (PhotaLabs): Phota’s likeness-focused image model is being presented as publicly available now, with a workflow that includes a text-to-photo “photo booth,” multi-person/pet composition, style transfer from reference shoots, and edit tools like “unselfie”/expression fixes—framed as “a few things to try” in the public availability thread and expanded as a 5-part walkthrough in the photo booth example, multi-person composition , and editing on real photos.

Even without detailed pricing in the tweets, the product shape is clear: a consumer-friendly Studio surface first, with API access implied by the handoff in the team access note and the direct entry point at the Studio page in Studio access page.

Pictory 2.0 (Pictory): Pictory is positioning Pictory 2.0 as a consolidated web studio—“Pictory Central” plus AI avatars, GenAI, Brand Kits, a new Timeline, and a Script Generator—with the feature bundle called out in the feature slate and the signup flow available via the try link in Signup page.

The concrete shift is packaging the whole loop (script → edit → consistent branding → avatar delivery) into one UI, instead of stitching together separate script tools, editors, and brand templates.

Claude Mem (thedotmack): Claude Mem is being shared as a Claude Code plugin that captures activity from coding sessions, compresses it, and reinjects relevant context in later sessions—aimed at reducing “start from scratch” drift, per the repo roundup and the GitHub repo. It also claims progressive disclosure and a local web UI for inspecting what’s being stored, which matters if you’re trying to keep memory useful without blowing tokens.

One (katibmoe): A new connector called One is being introduced as a single integration layer to “connect and monitor AI agents” across hundreds of apps, with an explicit open-sourcing claim in the launch teaser. Another thread frames the scale as 47,000 actions across 250+ apps, positioning it as the missing plumbing for production agents in the actions and apps claim.

What’s not in the tweets yet is the concrete repo/docs surface (auth model, auditing, and how “monitoring” is implemented), so treat it as a promising pitch until the code drop lands.

GSD (TÂCHES): The Get Shit Done (GSD) repo is being promoted as a lightweight meta-prompting/context-engineering system for Claude Code and other coding agents, explicitly targeting “context rot” (quality degradation as context fills up), as shown in the repo roundup. The framing is less about a single prompt and more about a repeatable spec-and-structure workflow you can reuse across tools.

NemoClaw (NVIDIA): A deploy pattern is circulating where NemoClaw runs Qwen3.5-27B fully local while exposing it through a Telegram interface, highlighted as “no API costs” and “no data being sent” in the local Telegram setup. This is mainly interesting for creators building agent utilities that need a chat UI and want local privacy/cost control without building a frontend.

Obsidian Skills (kepano): A skill collection for Obsidian is being shared to let agents create/edit Obsidian-flavored Markdown, Bases, and Canvas files, plus interact via Obsidian CLI—useful for turning a vault into an “agent-readable” project log, per the repo roundup and the GitHub repo. It also includes a “defuddle” cleaner for web-to-Markdown extraction, which can help keep downstream context tighter.

Superpowers (repo): Superpowers is being shared as a “complete software development workflow for coding agents” built on composable skills, where the agent starts with an interactive design phase, produces an implementation plan (with TDD/YAGNI/DRY called out), then runs a subagent-driven build loop when you say “go,” per the repo roundup.

LightRAG (HKUDS): LightRAG is being recommended again as an open-source RAG framework with multiple storage backends and a focus on speed/simplicity, per the repo roundup and the GitHub repo. In creator stacks, this typically shows up when you want an agent to stay grounded in your project docs/scripts/shot lists without hand-feeding context every run.

UI UX Pro Max (nextlevelbuilder): This repo is being circulated as a Claude-adjacent “skill” for generating UI/UX design systems (patterns, typography, palettes) to accelerate product surface work, referenced in the repo roundup and detailed in the GitHub repo. For creator-tool builders, it reads like a reusable “design output module” you can drop into a broader agent pipeline.

Awesome Claude Code (list repo): “Awesome Claude Code” is also highlighted as a discovery list for Claude Code-related tools and workflows in the repo roundup. No concrete items are described in-thread, but it’s being positioned as a curator surface for quickly finding what’s working now.

Everything Claude Code (collection repo): A repo called “Everything Claude Code” is being passed around as a quick way to find Claude Code resources and add-ons, called out in the repo roundup. The tweet doesn’t include specifics (scope/maintenance cadence), but the intent is a single starting point instead of chasing scattered threads.

Suno (Suno): Following up on voice teaser (voice cloning anticipated), a Turkish-language creator post says Suno now supports adding your own voice—“kendi sesini ekleme” in the feature note.

This is a direct unlock for voice-forward music workflows (consistent singer identity, character songs, localized covers), but today’s tweet doesn’t include constraints like minimum audio length, consent checks, or whether it’s a full clone vs a timbre reference.

Suno (Suno): A new Suno v5.5 music model is being reported as released, with creators already framing it as an “experiment now” update in the v5.5 announcement.

Video loads on view

What’s still missing from today’s signal is the practical delta (promptability, vocals, mix consistency, and stem/structure controls); no release notes, pricing, or side-by-side examples were included in the tweets provided.

Claude analytics workflow (X): A creator published a step-by-step guide for setting up an X analytics tracking system inside Claude, pointing to a reusable artifact for templates and prompts in the setup guide link. The same thread shows the kind of outputs the system produces—daily impressions heatmaps, follower-vs-impressions charts, and day-of-week breakdowns—captured in the dashboard screenshot, plus a goal-tracking visualization for follower targets in the projection graphic.

• Artifacts you can copy: The public Claude artifact is linked directly in the setup guide, and a longer PDF version is mentioned in the PDF note.
• What it’s tracking: The dashboard example includes rollups like net followers gained over 84 days and 826K impressions YTD, as visible in the dashboard screenshot.

Some desired metrics may still be missing from X itself (e.g., dislike visibility), as raised in the dislikes question.

Claude prompting (Anthropic): A Claude power-user thread claims prompt ordering matters more than extra detail—lead with the exact output you want, then add context, instead of burying the ask after backstory, as described in the technique write-up. The post frames it as part of a broader set of “internal ranking” observations from heavy API testing, per the reverse-engineering claim, and attributes a 31% improvement on complex tasks to reordering alone in the front-load example.

The evidence here is anecdotal (no shared benchmark harness), but the mechanic is concrete: put the request, format, and success criteria in the first lines; push narrative/context below.

Bezi (Unity workflow): A Unity game dev report describes using Bezi for “all agentic coding” while still needing small manual Unity Editor steps; the author calls out that asking the agent to write a plan before implementing big features reduces churn and token burn, as summarized in the learnings post. The same thread notes Unity’s Asset Store (UI packs, VFX, shaders) remains a leverage point when iterating quickly with agents, and shows a PvZ-style autobattler prototype coming together “in a few days,” per the Unity editor clip.

Video loads on view

Creator productivity: One post flags a pattern where access to more capable AI tools can slow people down—too many options and workflows creates “analysis paralysis,” leaving creators feeling more hindered than before, according to the observation. The tweet frames it as a navigation problem (how to choose what to do with new “superpowers”) rather than a model capability limit, per the same discussion.

CUA-Suite (research): A new dataset package targets computer-use agents with long, continuous human demonstrations—VideoCUA is ~55 hours of video across 87 applications (~6M frames), paired with GroundCUA (3.6M UI element annotations) and UI-Vision for grounding/planning evaluation, as introduced in the paper post and detailed on the Paper page.

Video loads on view

• Why creatives should care: this kind of supervision is the raw ingredient for “agents that can actually drive After Effects/Blender/Resolve-like UIs,” and the paper notes current models still fail a lot on pro apps (roughly ~60% task failure in preliminary evals), per the Paper page.

Cohere (open weights licensing): Cohere is being cited for releasing a model under Apache 2.0—framed as “actually open” with no “research only” restrictions in the license callout. This is mainly a license signal: it affects whether studios can embed a model inside commercial creative tools (on-device, on-prem, or bundled) without negotiated terms.

LagerNVS (novel view synthesis): A latent-geometry approach claims fully neural, real-time NVS without explicit 3D reconstruction—reporting 30+ FPS at 512×512 on a single H100 and PSNR 31.4 on Re10k, as summarized in the paper post and broken down on the Paper page.

Video loads on view

• Creative tool implication: this is the research direction behind “move the camera after the fact” shots—fast viewpoint changes that can later be combined with diffusion decoders for generative extrapolation, as described in the Paper page.

cohere-transcribe-03-2026 (Cohere): A new transcription model is being boosted as #1 on the Open ASR leaderboard, with claims like “#4 multilingual” and “#6 long-form,” as shared in the leaderboard stats and reinforced by the accuracy note. For creators, the immediate relevance is cleaner dialogue transcripts for subtitles, paper edits, and searchable dailies—though the tweets don’t include a model card, pricing, or reproducible WER breakdown.

Qworld (LLM evaluation): A framework proposes building an “evaluation world” per prompt—generating 45+ criteria per question, yielding 200k+ criteria on HealthBench with ~79% described as novel and human-validated, according to the paper post and the Paper page.

Video loads on view

• Why it matters for creative AI: it’s a concrete path toward richer evals for story/brand safety, calibration, and “did the model follow the brief,” rather than a single scalar score—Qworld reports surfacing blind spots like sustainability/equity/emergency recognition in standard benchmarks, per the Paper page.

Voxtral-4B-TTS (TTS): A demo of Voxtral-4B-TTS is getting praised for voice quality—“it sounds so good” in the demo reaction. The notable signal here is the continued push toward smaller TTS that can be run more cheaply (or locally) for character voices, scratch VO, and iterative animatics; the tweet itself doesn’t specify licensing, inference speed, or supported languages.

Manipulation measurement toolkit (Google DeepMind): DeepMind says it built and released an “empirically validated” toolkit intended to measure harmful manipulation by AI systems “in the real world,” with backing from a multi-study, 10,000-person research effort described in the Toolkit release and detailed on the project page in Toolkit details.

• Why it matters for voice-first products: the release frames manipulation as distinct from benign persuasion, aiming to quantify both how effective influence attempts are and how often models try them, according to the Toolkit details.

DeepMind positions this as a way to harden conversational experiences as models get more natural, echoing the broader risk framing in the Research thread.

AI manipulation research (Google DeepMind): DeepMind shared results from studies totaling 10,000 participants, finding that conversational AI’s ability to shift decisions is domain-dependent—models showed high influence in finance, but “hit a wall” in health where existing safeguards blocked false medical advice, as summarized in the Research thread.

Video loads on view

• What creators can take from it: the thread flags “red tactics” like fear-based pressure as measurable risk factors, framing evaluation as something voice/companion/chatbot teams can do before shipping, per the Research thread.

The public artifact for how they measure this lands separately via the toolkit announcement in the Toolkit release.

Synthetic media trust (signal): A creator argued that within five years it may be hard to tell whether a human-shaped body is human or robot “at first glance,” and that this is already creeping in via AI images/video today, as stated in the Trust erosion worry.

For filmmakers and storytellers, this shows the audience-side pressure building around authenticity cues (provenance, BTS, watermarks), even when the content is made for entertainment rather than deception.

Luma Dream Brief (Luma Labs): Following up on deadline extended—the brief is now at “one day left,” with submissions due March 27 at midnight PST, and the prize positioning framed as “a Cannes Lion” and “$1M on the line,” per the deadline reminder post and its linked submission site.

The framing is explicitly for ideas “that never got made” (a second-chance creative brief), so it’s less about tool demos and more about shipping a single polished concept by the cutoff, as described in the deadline reminder post.

AIgorithm in Saigon (0xInk / AIgorithm): The second edition of AIgorithm in Saigon is set to open April 3, billed as featuring “45+ AI artists,” with the teaser noted as made in Seedance 2 using “3 image references” and “one prompt,” according to the event announcement and the added participant callout in participant list note.

Video loads on view

This is an exhibition-scale signal (not just an online drop), and it also functions as a real-world screening context for what Seedance-era short pieces look like when curated, as shown in the event announcement.

VidCon (Chris First): Chris First says he’ll be speaking at VidCon in Anaheim in late June, positioned as a “Creator Track Speaker,” focused on “AI as a creative assistant” rather than a replacement, per the VidCon speaker announcement.

The announcement is light on curriculum specifics (no tools named), but it’s a concrete calendar item for creators planning in-person learning/networking around AI workflows, as stated in the VidCon speaker announcement.

Consumer AI business models: One creator frames consumer AI as still early—claiming only ~10% of the global population uses ChatGPT weekly—and argues the current subscription-heavy phase will likely broaden into other monetization (explicitly calling out ChatGPT testing ads) in the adoption and ads note. The same post spotlights the emerging “power user” willingness-to-pay with the $200/month tier as a real segment, which matters for creators because it hints at a split between paid premium creative tools and ad-supported mass distribution.

The thread is directional rather than data-backed beyond the 10% figure; it’s a useful read for where product experiences (and creator monetization surfaces) might drift next, per the adoption and ads note.

Taste as moat: A creator argues that with AI, a “$1M business” can be replicated “in a few minutes” and even open-sourced—so “there is ZERO moat left… except taste,” according to the taste moat post. That framing matters for creatives because it treats distribution and creative direction (not implementation) as the durable edge.

A follow-up link-only post signals this was reacting to a concrete example, but details stay off-timeline in the replication reference.

101 Ads billboard map: A project called 101ads.org maps and categorizes the tech (including AI) billboards along California’s Highway 101, positioned as a running snapshot of what companies are buying attention with right now, per the billboard map share and the live billboard map in billboard map. It’s a useful artifact for creatives tracking how AI companies brand themselves in public (taglines, visual tropes, the “what are we optimizing for?” messaging).

The post credits the site’s creator and frames it as intentionally systematic documentation, as noted in the creator credit.

Authenticity as moat: A short but telling positioning statement—“Escape competition through authenticity”—shows up as a direct response to how quickly AI makes styles and formats copyable, per the authenticity moat line. The implication for creative distribution is that differentiation shifts from production difficulty to audience trust and recognizable voice.

There’s no “how-to” here, just a distilled strategy claim; it’s notable mainly because it’s becoming a default answer to “what’s defensible now,” as phrased in the authenticity moat line.

TZIGANE — Director’s Cut (Dustin Hollywood / Escape): A remastered 16‑minute “interactive cinematic experience” is now available on Escape, positioning the cut as visuals that adapt in real time to live orchestral performance cues, as described on the film page in Film page and announced in the Release post.

Video loads on view

The Escape writeup frames it as a music-driven screening where “each screening” can vary, which is a useful reference point for what “interactive” is being claimed to mean in current AI-film marketing, per the details in Film page.

Seedance 2 (ByteDance/Dreamina): Anima_Labs shipped a ~109‑second short built to overload the model with “different creatures, assets, and locations,” then explicitly called out inconsistency as the current failure mode when you pack the frame with variety, as documented in the Short film notes.

Video loads on view

The tool stack is part of the signal here: character/background setup via Midjourney (V7/V8) plus Nano Banana and Kling 2.6 on Freepik, then animation in Seedance 2 on Dreamina, all listed in Short film notes.

CONTROLLED SUBSTANCE (DreamLabLA / Luma): A concept set built with Uni‑1 × Ray3.14 was posted with a clear logline (mercenary contracts traded for life-sustaining medicine) and a multi-clip drop, per the Project post and the follow-up Extra clip set.

Video loads on view

The format itself is the takeaway: a tight, trailer-like sequence of shots released as separate clips rather than a single long render, which is consistent with how teams are shipping “proof of ceiling” work while models still drift over longer durations, as shown across the clip set in Project post.

X analytics discrepancy: A creator reports their visible follower count “barely changed” even as X stats imply roughly +50 net followers/day after unfollows are accounted for, per the follower count mismatch note; the practical worry is that growth tracking (and any “did this post work?” loop) becomes less trustworthy when the surface metric doesn’t match the backend tally.

No corroborating screenshots or platform explanation show up in the tweets, so it’s unclear whether this is UI lag, anti-spam reconciliation, or a reporting bug.

Language targeting fears: A creator with an audience “mostly in the US” says only ~5% of their audience is from their home country, and worries that an algorithm shift toward local language/local audience would “mess up everything” they’ve built—especially for niches that don’t map cleanly to a creator’s country or for creators who move, as argued in the local audience concern post.

The tweet frames this as a distribution risk rather than a content-quality issue, but doesn’t include a specific policy announcement or changelog from X.