Skip to content

The Hidden Reason AI Advanced So Fast

2026-03-21 · other · Adil · recording, 12:22 · watch on YouTube

Topic: Studio work & business (also: Failure modes & fixes, Video generation & shot direction, Audio, voice & music, Prompting & AI assistants) · mode: Explainer

In this video history of AI, Adil traces the field from the Turing Test and symbolic AI to the moment computers were given data instead of rules, which he calls the core of modern AI. He moves through GANs, the Transformer and ChatGPT, which he calls AI's iPhone moment, then diffusion models, early video hacks and Higgsfield Soul. The story ends on the AI video race: Seedance's multi-shot storytelling with characters locked across cuts, Kling 3.0 and Seedance 2.0 released a week apart, and the AGI question.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — Introduction: from recognising a cat to Hollywood's panic

  • ★★ Adil says that in 1950 Alan Turing set aside the philosophy of consciousness and proposed a practical test of machine intelligence, the Turing Test. — 00:00:00

00:00:18 — The origin: the Turing Test, Dartmouth and symbolic AI

  • ★★ Adil says a group of scientists coined the term "artificial intelligence" at a 1956 summer workshop at Dartmouth, starting AI as a formal research agenda. — 00:00:28
  • ★★ Adil says that for decades AI was symbolic: rigid, hand-coded if-this-then-that rules, strong in maths but unable to tell a cat from a toaster in a picture. — 00:00:28
  • ★★★ Adil says the modern breakthrough came when computers were given data instead of rules: at its core, modern AI is pattern recognition. — 00:00:57

00:01:10 — The shift: pattern recognition, AlexNet and GPUs

  • ★★ Adil says the turning point for visual AI came in 2012 with AlexNet, a convolutional neural network that learned to recognise dogs, cars and stop signs from millions of labelled images. — 00:01:24
  • ★★ Adil says GPUs built to render video games turned out to be the right engines for AI's massive parallel computation, and that without gamers the AI revolution might never have happened. (hedged) — 00:01:48

00:02:05 — GANs and the origin of generative AI; the Transformer

  • ★★ Adil draws the line between traditional AI, which analyses data to make a decision, and generative AI, which uses that analysis to create new original data. — 00:02:15
  • ★★ Adil says Ian Goodfellow invented GANs in 2014: two AIs, a forger painting fakes and a detective spotting them, compete until the detective cannot tell the difference. — 00:02:15
  • ★★ Adil says the Transformer architecture of 2017 introduced self-attention, letting a model read a whole paragraph at once instead of word by word. — 00:02:40
  • ★★★ Adil says OpenAI released ChatGPT in November 2022, a large language model trained on massive internet data, and calls it the iPhone moment for AI. — 00:03:31 · ChatGPT, OpenAI

00:03:32 — LLMs: from Transformers to ChatGPT

  • ★★ Adil says ChatGPT proved that with enough data and computing power intelligence starts to emerge, and that the same Transformer architecture was then used to make AI see and move. — 00:04:01 · ChatGPT
  • ★★ Adil explains diffusion models: a model is trained to reverse noise added to real images step by step, then generates a new image by denoising pure noise. — 00:04:01

00:04:22 — Diffusion models and Stable Diffusion 1.5

  • ★★ Adil says Stable Diffusion 1.5 was the workhorse of the open-source community: for the first time anyone with a decent graphics card could generate high-quality images at home. — 00:04:48

00:05:06 — Deforum, AnimateDiff, ModelScope and spaghetti-eating videos

  • ★★ Adil says Deforum was a hack rather than true video generation: it zoomed and panned through Stable Diffusion's latent space, dragging one image into the next. — 00:05:17
  • ★★ Adil says AnimateDiff added a motion module that snaps onto existing image models and learned how people move, hair flows and light changes across frames. — 00:05:44
  • ★★ Adil says the first native video models, such as ModelScope with its spaghetti-eating videos, were trippy and heavily inconsistent but gave the first glimpse of AI motion. — 00:06:13

00:06:48 — Flux.1 Dev, Higgsfield Soul, Soul ID and Nano Banana; deepfakes

  • ★★ Adil says Flux.1 Dev in 2024 let AI follow complex spatial prompts and render legible, high-fidelity text on signs, shirts and screens. — 00:06:48
  • ★★ Adil says Higgsfield Soul moved image generation from random results to creative direction, specifically for fashion and aesthetics. — 00:07:11 · Soul, Higgsfield AI
  • ★★ Adil says Higgsfield solved character consistency with Soul ID: users upload a few photos and the AI learns the face as a permanent character to place in any scene. — 00:07:11 · Soul ID
  • ★★ Adil says models like Nano Banana turned image generation into a conversation: specific parts of an image can be edited and refined iteratively, as with a human designer. — 00:07:39 · Nano Banana
  • ★★ Adil says that with deepfakes nearly indistinguishable from reality, visual truth has become less reliable. — 00:07:39
  • ★★ Adil says video is 24 pictures per second and the model has to remember frame one for frame 24 to make sense, which is called temporal consistency. — 00:08:08
  • ★★ Adil says OpenAI announced Sora in early 2024. — 00:08:08 · Sora, OpenAI

00:08:24 — AI video race: Sora, Kling, MiniMax and Pika; keyframing and Veo 3.1 audio

  • ★★ Adil says Sora was the first model that truly understood 3D physics, but public access came a year later, when the hype had died down. — 00:08:37 · Sora
  • ★★ Adil says competitors such as Kling, MiniMax and Pika followed Sora, but for a long time brought only minor improvements and no big breakthrough. — 00:08:37 · Kling, MiniMax, Pika, Sora
  • ★★ Adil says Luma Dream Machine and Runway pioneered keyframing: upload a start and an end frame and the model generates the transition between them. — 00:09:07 · Runway, keyframe
  • ★★ Adil says Veo 3.1 combined start-and-end-frame interpolation with native, synchronized audio, adding sound to the motion it generates. — 00:09:07 · Veo 3.1, native audio
  • ★★★ Adil says ByteDance's Seedance introduced true multi-shot storytelling: creators feed images, video and audio at once and lock characters with tags so identity holds across cuts. — 00:09:35 · ByteDance, Seedance, multi-shot prompt

00:09:40 — Seedance multi-shot storytelling, Kling 3.0 and Seedance 2.0, studios and the law

  • ★★★ Adil says Kling 3.0 and Seedance 2.0 came out one week apart in 2026, with videos that behave like physical reality. — 00:10:03 · Kling 3.0, Seedance 2.0
  • ★ Adil says 2026 video models deliver minutes of sustained high-quality narrative where the illusion never breaks. (overstated) — 00:10:31
  • ★★ Adil says increasingly sophisticated AI video has drawn major studios' attention and raised unanswered legal questions, and that today's court and boardroom debates will define AI creativity. — 00:10:31

  • ★★ Adil says the law cannot keep pace with AI: laws are written in years, AI evolves in weeks. — 00:10:57

    “Laws are written in years. AI evolves in weeks.”

00:11:15 — The AGI question

  • ★★ Adil describes AGI as the field's holy grail: a machine able to understand, learn and apply its intelligence to any intellectual task a human can. — 00:11:24
  • ★ Adil closes by asking what will remain uniquely human if a computer can one day do everything people do, faster, cheaper and perfectly. (hypothetical) — 00:11:53

In the description and comments

  • ★★ Higgsfield says it has been at the centre of the generative shift and that Soul and Cinema Studio exist because of the technological lineage the video traces. — video description · Soul, Cinema Studio, Higgsfield AI
  • ★★ Higgsfield describes the video as the history of AI from Alan Turing's 1950 thought experiment to Hollywood studios suing AI companies in 2026. — video description

Mentioned

People: Alan Turing · Ian Goodfellow

Organizations: ByteDance · MiniMax · OpenAI · Runway

Concepts: artificial general intelligence · deepfake · diffusion model · generative adversarial network · keyframe · multi-shot prompt · native audio · temporal consistency · Transformer · Turing Test

Products: AlexNet · AnimateDiff · ChatGPT · Deforum · Flux.1 Dev · Kling · Kling 3.0 · Luma Dream Machine · ModelScope · Nano Banana · Pika · Seedance · Seedance 2.0 · Sora · Soul · Soul ID · Stable Diffusion 1.5 · Veo 3.1

Only in the description or credits: Cinema Studio

Replied to by

← Previous: Higgsfield Original Series: The First Complete AI Streaming Platform · Next → Higgsfield Chat: A Social Network for Collaborative AI Content Creation