Building the AWS of AI agentic commerce at https://nitter.cf/t.co/aCC3mYRz4z
Seattle, WA
Joined August 2012
- Tweets1.8K
- Following1.2K
- Followers188
- Likes1.9K
Yankz Teh ✨ retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Yankz Teh ✨ retweeted
We're open sourcing SPACE, the sandbox platform that runs every task in Perplexity Computer.
SPACE gives each agent task its own Firecracker microVM and controls credentials and outbound traffic from outside the VM.
Available now under Apache 2.0: github.com/perplexityai/SPAC…
Yankz Teh ✨ retweeted
Use voice chat, follow agent progress, and keep parallel work organized in the refreshed Codex CLI:
learn.chatgpt.com/docs/codex…
𝗪𝗵𝘆 𝘄𝗲 𝗱𝗼𝗻’𝘁 𝗽𝘂𝘁 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗮𝗴𝗲𝗻𝘁𝘀 𝗼𝗻 𝗢𝗽𝗲𝗻𝗔𝗜’𝘀 𝗔𝗴𝗲𝗻𝘁𝘀 𝗔𝗣𝗜
We use the Responses API and our own loop. We do not run production agents on the hosted Agents API.
𝗪𝗵𝗮𝘁 𝘄𝗲’𝗿𝗲 𝗿𝗲𝗳𝘂𝘀𝗶𝗻𝗴
Agents API is not another model endpoint. It is OpenAI running the Codex harness for you: sessions, compaction, recovery, subagents, tool orchestration. Even with a self-hosted sandbox, they still own the control plane. Compute location is not the same as ownership of the agent.
𝗧𝗵𝗲 𝘁𝗿𝗮𝗱𝗲
Teams that adopt Agents API will ship faster this quarter. That is real. They skip harness work we still have to maintain. We accept that lag on purpose.
𝗪𝗵𝘆 𝘀𝗽𝗲𝗲𝗱 𝘁𝗼𝗱𝗮𝘆 𝗶𝘀 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲
They see the workflow.
The loop, tool sequence, failure modes, and what “done” looks like all live on their side. That is the training distribution for the next product they ship into the same category. Lock-in is the product.
Session state, compaction policy, and recovery are not portable. Leaving means rebuilding the agent, not swapping a model string. Assistants and Agent Builder already showed how these objects get deprecated while the dependency stays.
Self-hosted sandbox does not fix this.
Your executor still takes commands from their orchestrator and sends results back. Data residency is US-only. Zero Data Retention is not available. Compliance and information boundary stay theirs.
The moat moves to the runtime.
If the model commoditizes, the company that owns the harness and the traces owns the vertical. In two or three years the likely outcomes for a hosted-loop shop are high switching cost, or a thinner wrapper competing with OpenAI’s own packaged agent for the same job.
𝗪𝗵𝗮𝘁 𝘄𝗲 𝗸𝗲𝗲𝗽 𝗶𝗻𝘀𝘁𝗲𝗮𝗱
Model calls and hosted tools via Responses. Orchestration, state, approvals, audit trail, and sandbox in our process. Multi-provider from day one. The workflow (categorization, reconciliation, dispatch, invoicing) is the company. We do not put that loop on a vendor’s servers.
𝗥𝘂𝗹𝗲
Prototype on whatever is fastest.
Production agents that encode our playbooks do not run on a managed harness we don’t control.
Speed we don’t own is not an advantage.
Yankz Teh ✨ retweeted
🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.
🔹 Introducing the smallest model in our new architecture family, with native visual understanding.
🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models.
1/6
Yankz Teh ✨ retweeted
AI investments stall because they lack executive support and a strong technological foundation. mitsloan.mit.edu/ideas-made-…
Yankz Teh ✨ retweeted
we officially hit the $0 cost per task era.
intelligence is commoditizing way faster than anyone expected.
when OpenCode + Gemini3.7 flash (high) is sitting at $0.00 and DeepSeek is $0.07, burning $11 per task on frontier models in production is financial negligence for 90% of dev workflows.
benchmark quality gaps don’t matter when the cost delta is literally 10,000%.
the math gets wild when you check raw token pricing:
- Gemini 3.7 Flash: $0.75 in / $3.75 out
- DeepSeek V4 Flash: $0.14 in / $0.28 out (cache miss)
(deepseek $0.0028/m input tokens with cache hit)
for Gemini to land at $0.00 vs DeepSeek’s $0.07, either Google's prompt caching hit rate + token efficiency is completely cracked, or AA hit a rounding artifact.
what’s your daily driver stack right now?
Yankz Teh ✨ retweeted
Today we're introducing Gemini 3.7 Flash, our most intelligent workhorse model yet for coding and agents.
This model brings substantial gains across software engineering, web development, and complex knowledge work.
Now through the end of the year, Gemini 3.7 Flash is available at an introductory price* of $0.75/1M input and $3.75/1M output tokens — making it easier and more cost-effective for developers and customers to scale production-ready agents.
Yankz Teh ✨ retweeted
My skills repo has 160K stars, 7.5m downloads...
...and no tutorial.
So, here it is. Watch me walk through the essential skills:
- /grill-with-docs
- /to-spec
- /to-tickets
- /implement
- /code-review
It's the whole flow, end-to-end. Enjoy:
Yankz Teh ✨ retweeted
Big glass of cold water thrown in my face.
Found an error in our financial projections late last night. Accountant was offline. Asked Claude to model the impact on revenue, cash, and EBITDA.
It gave me a catastrophic answer. And was so damn confident about it.
I legit spiraled for 2-3 hours, even got @thesamparr involved (stupidly).
Accountant finally called. Claude was completely wrong.
When I showed it the mistake, it apologized and said it would never do that again.
I lost a few years off my life, and needed Sam to talk me off a ledge, both unnecessarily.
Good reminder, even opus 4.8 has a long way to go…
And having a co-founder to balance out your panic attacks is priceless.
Make me feel like less of an idiot and share your ai disaster stories please. 🙏
