Mission #20: Swap the model, keep the pipeline
Models change under you. Last week OpenAI launched GPT-6.1 Sol at one-fifth the token price of GPT-6 Astra. Next month it will be another lab. If your pipeline has a model name and a provider call baked into code, each of those is a rewrite, a test run and a redeploy.
On RocketRide Cloud, the model is a setting. Every LLM node picks a profile, a preset that carries the model id and the token limits. The OpenAI node defaults to GPT-5.2; GPT-5.4, GPT-5.4-mini and GPT-5.4 Pro sit in the same list. You add the credentials, usually an API key kept as a RocketRide environment variable, and nothing else. Change one setting and the pipeline runs on a new model.
Switching providers is nearly as small. All 17 LLM nodes in the catalog, one per provider, OpenAI, Anthropic, Gemini, Bedrock, Mistral, Ollama and the rest, take questions in and hand answers out on the same lanes. Swap the node, pick its profile, and the nodes on either side never notice. When a model lands before its preset, as GPT-6.1 Sol did, the custom profile takes the model id and key directly.
The model should be the easiest thing in your pipeline to change. Build it so it is.
How profiles work
A profile is a named preset inside a component. In the editor you pick it from a list: GPT-5.2, GPT-5.4, GPT-5.4 Pro. In the .pipe file it is one line, "profile": "openai-5-4", which pulls in the model id and the token limits. The section named after the profile is optional and usually just the credentials: an apikey for most providers, an access key, secret key and region for Bedrock, a server URL and no key for Ollama. Leave the model id and token limits alone unless you are on custom.
Two things to know before you swap. Embeddings are profiles too, miniLM, miniAll and mpnet, and they run server side with no key, but ingestion and search must use the same one, because the vectors have to match. And a pipeline that uses RocketRide environment variable placeholders carries only the reference, never the key; a key pasted straight into the config is stored with the pipeline, so use the placeholder.
Two rooms left this week, then the report from the rooftop.
- Zero Human Startup Hackathon, today, Wednesday 7 October, 10:00 AM to 6:00 PM Pacific at Sports Basement on Bryant Street, San Francisco. A build day for a startup run by three to five AI agents, product, research, GTM, sales, working together through real workflows, with judging at 4:00 PM. Phase 2 continues until Sunday 11 October, and both phases carry prizes. Mansi More hosts alongside the organizers. Free, approval required. Request a spot on Luma.
- RocketRide Office Hours: Approval Gate or Spend Cap?, Thursday 8 October, 10:00 to 11:00 AM Pacific on Discord. Part 3 of the agentic building series: should the control on an agent be a human who signs off, or a boundary it cannot cross? Approval prompts that stop getting read, caps beyond spend, and which actions are reversible enough to let run. With Mithilesh Gaurihar, Josh Phillips, Krish Garg and Ansh Kaushik. Details on Luma.
Last week, for the record. Monday night Builders x Bosses x Beers wrapped Day 1 of #SFTechWeek on the rooftop at Club 345: short talks from teams actually shipping, drinks until nine, the AI Barmen mixing them.
Tuesday at Cloudflare, Josh Phillips ran Zero to Agent: 90 minutes to build a claims desk for Specter Mutual, a ghost insurer with unhinged claims, that cannot be sweet-talked out of the rules, then the room turned fraudster and tried to break it. The calendar has LA Tech Week next week.
The week's AI news that changes how you build.
At DevDay on 29 September, OpenAI launched GPT-6.1 Sol at one-fifth the standard input and output token prices of GPT-6 Astra, the model it decided not to ship after its own researchers raised concerns. OpenAI says 6.1 Sol beats GPT-6 Sol on programming and debugging, document understanding and multistep workflows, and it is live for Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. Sam Altman also introduced Dots, an always-on agent he described as an AI helper that always has your back. For anyone running a pipeline, the lesson is in the price list: the model underneath you changed price and rank in one afternoon.
Voice agents got 20 more languages and a 10-second clone.
ElevenLabs released v4 and v4 Turbo on 28 September. Language coverage goes from 70 to more than 90, with Japanese, Brazilian Portuguese, Mandarin and Cantonese named as the biggest gains, and a voice can now be cloned from 10 seconds of audio. Expression is controlled with inline tags that stack, and for agents the audio starts as soon as the LLM starts producing an answer, which is the part that makes a phone call feel like a conversation. Both models are in the API and the agents platform.
AMD paid $8.2 billion for Fei-Fei Li's world models, and the chips will follow the workloads.
AMD agreed on 28 September to buy World Labs for $8.2 billion, with Fei-Fei Li joining as executive vice president and chief scientist. World Labs builds models that understand physical space; its first product, Marble, generates simulated environments for entertainment and for training robots. AMD says the frontier workloads it inherits will shape its chip roadmap, a direct answer to Nvidia's Cosmos world models. The deal is expected to close before the end of 2026, pending regulatory approval.
Member discussion