Ep 12: Why is AI not consistent? | Ft. Travis Good x Charlie Hu

Large models from GPT-5 to Anthropic Opus 4.5, 4.6, and 4.7 now generate professional level code, content, and risk reports as good as senior software engineering directors or research directors. But companies need consistency to use AI at institutional level. Agents store everything you send them in short-term memory, long-term memory, and company memory. Without proper workflow, they just YOLO and figure it out themselves, giving you random low quality slop instead of consistent high quality output. Single source of truth matters because agents need correct information from the same GitHub repository, not random data from other people. The problem is large models are probabilistic, not deterministic.Harnesses fighting harnesses creates real churn where closed source providers have their own guardrails, auto routing, and bespoke infrastructure constantly changing responses before you get them. Claude might suddenly refuse questions about financial reports or biomedical chemical symbols because it registers as a threat at Anthropic. Context engineering gets tricky with RAG, graphs, and RLM where agents do text searches and spin out chunks to sub agents that report back. Once agents are successful, you need to capture the workflow steps that worked and repeat them consistently. With employees, you can reprimand or fire them for poor behavior, but with agents it's difficult to pin blame because systems are so complex and non-deterministic. Ambient offers proof of what ran at what time and what was generated so you can trace exact context used and results produced. How do you hold non-deterministic AI agents accountable when closed providers hide their harnesses underneath? Ambient’s design centers on one large network model (600B+ parameters) and its fine-tunes, so miners optimize for utilization instead of wasting cycles across fragmented “model marketplaces.” The project positions verified inference as infrastructure: low overhead verification, high throughput, and a path toward open, transparent training over time.•••About COCO AI:Find COCO AI on X.•••About Charlie Hu.Charlie Hu is the co-founder of COCO AI, currently based in Singapore with seven years of building Web3 infrastructure. DeFi brought him back to Ethereum, and he joined Polygon in 2021 as head of APAC for one and a half years, driving metaverse and consumer-facing applications ecosystem building. For the last three years, he co-founded Bitlayer, one of the leading Bitcoin L2 networks, and built the BVM bridge and YBDC yield-bearing BTC protocols with significant success. With a large engineering team, he started using Cursor and Claude Code to reduce costs through AI, initially building enterprise AI solutions for recruitment before pivoting to multi-agent enterprise teams over the last six months.Find Charlie on X.•••Hosted by Travis Good, CEO & Co-Founder of ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Ambient⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠: an SVM-compatible PoW L1 that will serve as a cornerstone of the agentic economy. Ambient has raised $7.2M from a16z, CSX, Delphi Digital and Amber Group. It is 10x more efficient than incumbent crypto AI systems. Ambient believes AI is money and is working to democratize it. Find us on: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠X⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ | ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Youtube⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ | ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠LinkedIn⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ | ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Ambient Chat⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Find Travis on: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠X⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ | ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Linkedin⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠Find our Litepaper ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠here.