Back to blog
Sep 06, 2026
5 min read

Once it can write, teach it to see, speak, and film

Text is the foundation, but the same models now listen, talk, look at pictures, and generate images and video. The eighth stop on the AI roadmap: what each of those actually is, what works in production, and what still breaks.

The eighth stop on my AI roadmap is the one that opens up once text is solid. Everything before this was words in, words out. But the models are now multimodal: they take images and audio as input, and separate models generate speech, pictures, and video. Same ideas, more senses. This stop is a map of what each one is, what is genuinely production-ready in 2026, and what still falls apart.

Same assistant, more senses

Picture the assistant from earlier stops, the one who could only read and write. Now they can listen to a voice note, look at a photo you hand them, reply out loud, sketch something, and cut a short clip. Each of those is a different tool with different maturity, and the useful thing is knowing which is a commodity you can lean on and which is still a party trick.

flowchart LR
  T[Text model at the core] --> V[Vision: reads images]
  T --> S[Speech in: transcription]
  T --> O[Speech out: text to speech]
  T --> I[Image generation]
  T --> M[Video generation]

Seeing: vision is now a commodity

Reading images is the easy win. Every major model provider now accepts an image alongside your prompt through the same API you already use: hand it a screenshot, a receipt, a product photo, and ask about it. No separate service, no new skill. If your app has images and questions about them, this is a single extra field in a call you already make.

Listening and speaking: voice crossed the line

Two halves. Speech to text (transcription) is dominated by Whisper-style models, with real-time services for live use, and it is reliable enough for production. Text to speech has crossed the uncanny valley: the best current voices are hard to tell from a human recording.

Put them together with a text model and you get a voice agent, and the piece that makes it feel natural is a realtime audio API: low enough latency that a conversation flows instead of lurching. This is the pattern behind phone agents and voice assistants.

NOTE

Latency is the whole product in voice. A great answer that arrives two seconds late feels broken in a conversation. Stream the audio out as it generates, exactly as you stream text, and measure the time to first sound.

Drawing and filming: impressive, with sharp edges

Image generation is mature for the common cases: product mock-ups, illustrations, variations on a style. The open models let you run it yourself; the hosted ones are a single API call. It still stumbles on precise text inside an image and on very specific layouts.

Video generation is production-ready for short cinematic clips, and that qualifier matters. The current generation handles a few seconds of controlled camera and character work well, and still fails on hands, on legible text in the frame, and on keeping a character consistent across a long sequence. Treat it as a tool for short, curated pieces, not a replacement for a video pipeline.

WARNING

Generation output needs human review before it ships, every time. Hands with six fingers, text that is almost English, a logo that is nearly yours: these are common, and a model will not flag them. Review is not optional at this stop.

The tools you actually reach for

JobPopular toolsNote
Vision (image in)Any major provider’s modelSame API as text; add an image to the prompt
Speech to textWhisper (open), DeepgramWhisper is free/OSS; Deepgram for real-time
Text to speechElevenLabs, OpenAI TTSHosted, paid; free tiers to start
Realtime voiceOpenAI Realtime API, and equivalentsLow-latency conversational audio
Image generationFlux, Stable Diffusion (open), hosted APIsOpen models run locally; hosted pay per image
Video generationRunway, Kling, Google VeoShort clips; review output

More senses, same rules

The lesson from this stop is that the fundamentals from the text stops carry over completely. Latency still decides the experience, so stream. Output can still be confidently wrong, so review it before it reaches anyone. Cost still scales with volume, so budget. What changes is only the input and output format. Treat each new modality as the same model with a new sense, not a new kind of magic, and the earlier discipline keeps working.

New senses, old discipline.

Sources

Read next