Paddy Srinivasan

.@Kimi_Moonshot K3.1 Pro is #4 of 221 models on the @ArtificialAnlys Intelligence Index (as of Oct 23), and you can download the weights. It’s now on DigitalOcean Serverless Inference. Upgrading from K3 is a one-string change. do.co/4xsz0fI

If you’re building with agents, models or data, we’ll probably be in the same room this month. @Box|Works, @lancedb Reverie, @sgl_project Summit, @TomorrowXSummit, @foundersysk, plus a happy hour with @sgl_project after their Summit. Come find us.

Most agents ship once and stay the same. The best ones learn from every run. Nov 12 in SF: happy hour with @NVIDIAAI, @sgl_project, and @MLHacks after SGLang Summit. Lightning talks and demos on building that loop. Request your spot: do.co/48pUak2

Picking a good inference provider is difficult. Picking the best inference provider for an open model is surprisingly hard: The reason: no single provider does it all. On OpenRouter, GLM 5.3’s GPQA score ranges from 78% to 89% (quite a large gap, despite the caveats†), and the trends in GPQA is not strongly correlated with other tasks, so it varies by workload. Why discuss this now? Recently, Jon Barron and colleagues showed that standard softmax implementations hide two silent bugs: Softmax Collapse, where the target gradient rounds to zero, and Softmax Cancellation, where identical logits still produce a gradient. Widely speaking, inference has the same problem, except it's harder to get something agreed to be a 'correct' outcome. In my experience, providers on alternative chips tend to have worse silent framework bugs, though these issues are apparent on NVIDIA inference frameworks as well. MoE models are especially fragile: the router bias and router output need fp32, etc. (there's a long list of half-patched issues). As deployment engines grow more complex, these bugs get almost exponentially harder to catch. (Note†: SGLang or vLLM in their default configurations should give the best accuracy. The issue come from what providers add, like some custom quantization, speculative decoding, and custom kernels, or even simply forgetting to update to the latest versions of the frameworks that improve accuracy. And consider the case where you have no open reference at all (i.e they take care of both post-training and inference, so you don't own your own canon copy of the model, where you wouldn't know if there's some error in the stack silently reducing your quality).

A model’s suggestion to call a tool is a request, not authority. Action Gateway keeps credentials out of model context and checks every tool call against policy: allow, deny, or ask a human. It aligns with @nvidia’s Open Agent Safety Platform. do.co/4ierrTS

.@MiniMax_AI reports 73.67% on DeepSWE and 72.5% on AutomationBench-AA, which it calls the best result among open models. M3.1 Flash is now on DigitalOcean Serverless Inference and Dedicated Inference. $0.15 input, $0.60 output per 1M tokens. do.co/4xsz0fI

Who can help get a @bot app published ASAP?

.@deepseek_ai reports V4.1 Flash 1017 scores 77.0 on DeepSWE v1.1, up from 74.2 on V4.1 Flash, with shorter outputs and faster task completion. Now on DigitalOcean Serverless Inference and Dedicated Inference. Already on V4.1 Flash? Switch to deepseek-v4.1-flash-1017 and keep the same endpoint and Model Access Keys. do.co/4xsz0fI

hey! genuinely interested to get your feedback if you want to chat more about your experience with DigitalOcean. sincerely, someone from product who wants to make it better. DM me if you are interested in chatting more for account credit or some swag! 🫶

.@reflection_ai reports Beam matches leading open models on quality with about 3x fewer reasoning tokens. It's a 501B sparse MoE with only 23B active per token, and Reflection's first open-weight model. Now on DigitalOcean Serverless Inference and Dedicated Inference. $0.40 input, $1.20 output per 1M tokens. do.co/4xsz0fI

A hypothesis: agents don't need a sandbox. They need a platform that thinks, acts, remembers, and improves as one system. There's a debate running this week about sandboxes. One camp says sandboxes are dead and the future is a private cloud for every agent. Another says a sandbox is a compute primitive, like Postgres or Redis, and primitives don't win markets, platforms do. A third says sandboxes are a winner-takes-most market. I think all three are arguing about one layer of the problem. My hypothesis is that agents are the first workload that exercises every layer of a platform inside a single loop. They think, they act on real systems, they remember, and they improve. A web server touched compute and maybe a database. An agent touches inference, execution, tools, storage, retrieval, telemetry, and evaluation, in one task, often in less than a minute. So what an agent needs isn't a better sandbox. It's a platform where those four things run as one system. If that's right, the unit of competition isn't the sandbox and it isn't the model. It's the integrated cloud for agents. Whoever builds one that's compelling across agents, data, and model access at the same time wins. Let’s start with how an agent actually runs. One “prompt” fans out into a dozen model calls, a web search, a database query, a code execution, a tool call with a credential, a checkpoint, and a trace. Today a typical team buys those from five or six vendors and glues them together. Six vendors means fifteen seams. Every seam is an integration someone maintains, a credential someone stores, a network hop that adds latency, an egress fee, and a place where the trace dies. Point solutions are excellent but partial: most sandbox companies don't run inference, and model companies don't manage agent memory or observability. The loop is what breaks on a stitched stack. Models have to be routable, and routing only works when inference sits next to the agent. Every few weeks a new model tops a chart. Open-weight models went from <20% of token volume on our Inference Engine in April to >85% by August. My read: the model is the most swappable layer in the stack, so you want a router in front of it, not a hard-coded choice. We built an Inference Router for that. On Kimi K3 we see an aggregate cache-hit rate above 90% across our own workloads. The router has to know what the inference layer has cached, and the cache has to sit next to the sandbox that's going to need it. That's a function call on one platform. The agent and harness layer is where the business logic lives, so the runtime has to be built for it. I believe the harness, its prompts, its tool permissions, its memory, and its evals will hold more of a company's business logic than the model ever will. The model is rented. The agent is yours. So we built Managed Agents from the ground up as purpose-built primitives, not VMs with an agent sticker on them. Every session gets its own Firecracker microVM, resumes from pause in a few hundred milliseconds, and is billed for active CPU at $0.044 per vCPU-hour, so the wait costs zero. Action Gateway gives agents governed access to 20K+ tools across 600+ providers through one MCP endpoint, matched intent to tool with 99%+ accuracy in our testing, and brokers credentials at call time. The agent never sees the key. You can only make that promise when the runtime, the gateway, and the secret store are the same system. Data and recursive loops are what make an application more intelligent over time. This is the part I believe in most and the part with the least proof so far, because the products are weeks old. The hypothesis: two teams start with the same model and the same harness, and within a quarter they diverge, and the one with the loop wins. On DigitalOcean the loop runs on two services. Agent Data Services is one governed catalog where video, documents, logs, and agent sessions land agent-ready and searchable by meaning. Agent Learning Services closes the loop: Insights records every step, Signals flags the sessions where the agent stalled, looped, or lost the user, and Evaluations and Simulations test the fix before a customer sees it. In an internal demo, an optimized support agent went from failing to passing it with 90% fewer input tokens. Intelligence stops being something you rent via models and becomes something your application accumulates, tuned to the exact problem it's solving. And the whole thing has to go down to the floor. DigitalOcean builds the routing model and the microVM runtime operates the GPUs they run on, across 20 data centers with NVIDIA HGX B300 and AMD MI350X fleets. Much of the agent infrastructure market rents GPUs from someone who rents a data center from someone else, with margins stacked on margins. My view is that active CPU billing, a warm KV cache next to the sandbox, and a router that can see the inference layer are only possible when you control the data center, the hardware, and the software as one stack. The obvious objection is that "integrated" is a polite word for lock-in. I'd argue the opposite, and I think it's the real test of whether a platform is integrated or just bundled. Integration should be about removing seams, not closing doors. Bring Claude Code, Codex CLI, OpenCode, Hermes, LangGraph, or your own OCI image. Hit one endpoint compatible with OpenAI, Anthropic, and 60+ Open models. Bring your own key for a frontier model, or run open weights we host. Our routing model and the proxy under it are open source. The point is that the seams are gone, not that the exits are. One early signal that the hypothesis might be right: roughly 70% of our large AI customers have attached a one or more platform primitives. Agents come in through inference or a sandbox and pull the rest of the cloud behind them. That's what I'd expect if agents really do need every layer at once. So, back to the debate. The platform that wins won't be the one with the best sandbox, and I don't think it'll be the one with the best model. My bet is on the platform where an agent thinks, acts, remembers, and improves as one system. DM me if you are interested in trying our Agent primitives. #digitalocean #agents #Agenticloop #learning #Context #sandbox #Runtime #DOCN