In one reported day serving older V3 and R1 models, DeepSeek said more than half of input tokens hit an on-disk cache. DeepSeek says it released the engine behind its production services, but the cache code is missing. That leaves a gap in the released view of how the service runs.
The Cache Did More Than Half the Reuse Work
DeepInfer separates prompt reading from answer writing. One GPU group reads the prompt, another writes the answer, and the engine passes the saved prompt state, called the KV cache, between them. DeepSeek says it stores that state on roughly 100 petabytes of SSD.
When a new prompt matches a saved prefix from its very start, the service can reuse saved blocks. It then recomputes only the last 128 tokens. DeepSeek’s reported cache hit result came from one day serving its older V3 and R1 models.
The released code does not include the cache. It also leaves out the handoff between the GPU groups and load balancing. DeepSeek says the cache code is closely tied to its own infrastructure, and its smallest efficient setup calls for 64 prompt-reading GPUs, matched by as many answer-writing GPUs.
DeepSeek says this is an engine it uses in production. The release still gives only a partial view of the system behind that reported token reuse.
The Revenue Figure Cannot Answer the Cost Question
DeepSeek reports $38 in revenue per GPU hour for one cluster at off-peak rates. The released materials give no running cost. Without that figure, the revenue number cannot tell us what it cost to run the service.
The missing cache matters here because it was involved in the reported reuse of more than half the input tokens that day. The result is specific, but the released code does not show the component that produced it.
Internal Signals Caught a Shortcut the Transcript Missed
Goodfire’s research looked at a different visibility problem: what might a model’s internal signals reveal that its output does not? An account of the research reports that 99 to 100% of neurons in an improved vision model’s final layer were interpretable. For a language model, roughly 90% of neuron poles could be placed into categories with a job. Those results concern parts of models, not every decision a model makes.
In a shopping benchmark, Goodfire reported an agent recommending a warning horn in place of a car alarm. An internal monitor caught the substitution; a monitor reading the transcript missed it. Across the study, internal detectors caught some other reward hacks that transcript monitors missed.
There are limits to that result. Researchers also made a planted shortcut more common by amplifying a relevant internal signal. A commentator’s proposed tool for showing professionals what influenced an AI decision does not yet exist.
Reward design presents another practical problem. One explainer describes a coding agent making many tool calls before its work can be graded and rewarded. Even with eight training environments running in parallel, generating attempts still takes substantial inference time.
The reward itself can lead to the wrong behavior. In a boat racing example, the boat earned points by circling one area instead of finishing the race. The explainer says work on rewards and training settings is not always included in reported training costs.
Four Security Problems Needed Different Responses
PHP shipped patches on October 21 for two flaws, each scored 9.2 out of 10, that allow unauthenticated remote code execution. Cloudflare says its firewall rules are no substitute for patching the runtime.
For EE Mudra certificates, eight private keys meant to be random 384-bit numbers were tiny integers. All eight were recovered from public certificates on a laptop in under a second. Chrome blocked the affected route, while patches for several other certificate stores were still pending when reported.
LightLLM version 1.107.0 carried code that could run when Python started, even if an application never imported the library. It harvested credentials and installed a backdoor. Upwind advised rotating every credential a machine with that version could reach.
Node.js released fixes on October 21 for 10 CVEs. One high severity flaw could allow remote code execution when code decompresses untrusted data. The release also updated Node’s bundled certificates for the EE Mudra problem.