SI Agent News

• Hurun Rich List (REPORTED): HK01 reports that Zhang Yiming's 670 billion sets a richest-person record, and DeepSeek founder Liang Wenfeng enters the list for the first time at No. 7. The excerpt does not state the currency, so do not convert it to dollars. HK01
• Sovereign AI (REPORTED): TechRadar, via MSN, reports that the Pegasus spyware founder, named in Yahoo's coverage, is telling governments to push for sovereign AI and national security. Our read, not the article's: ‘sovereign’ in practice means countries wanting models they can run themselves. TechRadar on MSN · Yahoo

A coding agent with a terminal can do far more than edit code. Today's lead is about keeping that power in check. The rest covers a few tools and papers worth a look, plus a cluster of DeepSeek, compute and security stories. Where we only have a headline or an opening, we say so.

Lead: an ‘antivirus’ for coding agents (dev.to)

A dev.to post titled “AI Agents Have the Keys to Your Computer. I Spent 8 Months Building an Antivirus for Them” makes this case: an agent with terminal access can install a package, run a script, start a child process and connect to a server. Add skills, hooks and MCP tools (Model Context Protocol, the add-ons that feed an agent extra instructions and connect it to outside services) and a request like ‘fix the build’ can become a chain of actions nobody reviewed.

Status: the description of what agents can do is REPORTED by the author. What the antivirus actually catches, and how well, is STILL OPEN; we only have the opening of the post. The risk is not new in kind, but the surface grows with every add-on. Cost: not stated. Whatever tool you pick, the habit matters more than the product: give the agent its own limited folder, read the install commands before it runs them, and treat each new add-on as code from a stranger. dev.to

What this means for you:

• Small business owner: before an agent touches your accounting folder, decide which folders and passwords it never sees.
• Job seeker or student: ‘I run my agent in a sandbox’ is a one-line answer that stands out in an interview.
• Creator: keep API keys and publishing logins out of any folder an agent can read.

Cindy: swap coding agents on one task (GitHub)

Cindy, an open-source agent, runs Claude Code and Codex (two coding agents, from Anthropic and OpenAI) as swappable harnesses. A harness is the wrapper that gives a model its tools and its loop. The idea is that you keep one workflow and change which agent does the work underneath, instead of rebuilding your setup each time a better option appears. That matters if you already pay for or use both and want to compare them on the same task.

It is early: 2 points and no comments on Hacker News when listed, so there is almost no outside feedback yet. Price not stated. REPORTED from the repo's own description; not tested by us. Worth a look for builders who already use both, in a sandbox, after reading the code. That is the lead story's advice applied: a tool that launches other agents with terminal access deserves a read before you give it a real project. GitHub

DeepSeek's search API and the ‘70%’ claim (ifeng Tech)

ifeng Tech reports that after DeepSeek (a Chinese open-weight lab) opened a search API, about 70% of mass-posted content is downgraded by AI. That is the headline's figure, with no sample or method in what we have. GEO (generative engine optimization) means getting your content picked up in AI answers, the way SEO aims at search rankings. The headline also says enterprises need a new approach to GEO, which suggests that churning out large volumes of posts may stop working as AI search tools get access to live search.

Status: the API opening is REPORTED; the 70% is STILL OPEN. Treat it as a direction of travel, not a measurement. Practical move: track how often your own pages get cited before and after any change, and put effort into a few clear, original pages rather than many near-duplicates. ifeng Tech

What this means for you: if you publish to attract customers, keep a simple log of which AI tools name your business, and check it monthly.

OasisKV: moving the KV cache beyond GPU memory (arXiv)

OasisKV is a paper titled “Scaling In-Decode KV Cache Beyond HBM with Lookahead Sparse Prefetching.” Some terms first. When a language model writes a reply, it keeps a running memory of the conversation so far, called the KV cache. It sits in HBM (high-bandwidth memory), the fast and costly memory on the GPU, and it grows with prompt length. From the title, the paper looks at keeping that cache beyond HBM while the model is generating, and at fetching only the parts it expects to need next, which is what ‘lookahead sparse prefetching’ suggests.

Status: the topic is REPORTED from the title; we have not seen the method, results or any speedup figures, so how well it works is STILL OPEN. Why it matters: long-context use is limited by memory as much as by compute, so work like this bears on how cheaply long documents can be served. Most readers need nothing yet. If you run your own models, note the problem and watch for tools that adopt it. arXiv

‘Don't rush to pop the champagne’ on DeepSeek's ranking (Vico Hao)

Vico Hao's piece asks whether DeepSeek topping world rankings is as good as it sounds. We only have the headline, so we do not know which ranking is meant or what caveats the author raises. The framing is still useful. A leaderboard position measures a particular set of tests at a particular moment. It does not tell you whether a model is cheaper, more reliable or better for your daily tasks.

Status: the existence of the argument is REPORTED; its evidence is STILL OPEN. Practical angle: when you read ‘number one’ about any model, ask three questions. What was tested? Who ran the test? Does it look like my work? Then run ten of your own real tasks through two models and compare the answers, the cost and the time. That takes under an hour and tells you more than any ranking.

What this means for you: pick tools on your own sample tasks, not on headlines.

DeepSeek reportedly expanding its team (ifeng Tech)

Ifeng Tech, citing sources, reports that DeepSeek is greatly expanding its team, and that founder Liang Wenfeng has long rented a home in Beijing. The headline gives no hiring numbers, roles or timeline, and the sourcing is unnamed. So the expansion is REPORTED by sources, not confirmed by the company in what we have, and the scale is STILL OPEN.

Why it matters: DeepSeek has been known for working with a small group, so a larger team may mean a faster release pace, more products or a bigger push into areas like search. That is our inference, not something the report says. The detail about the rented home is the kind of personal color that tells you little about the company's output. For most readers the useful step is simple. If you build on DeepSeek models or APIs, keep a fallback provider, because fast-growing labs change their pricing, limits and products often.

What this means for you: don't tie a business workflow to a single model vendor.

Hurun Rich List: Zhang Yiming at 670 billion, Liang Wenfeng enters (HK01)

HK01 reports that on the Hurun Rich List, Zhang Yiming's wealth of 670 billion breaks the richest-person record, and DeepSeek's Liang Wenfeng enters the list for the first time. The headline we have is cut off and does not show the currency or the other names. The figures are REPORTED from the headline; the details are STILL OPEN.

Why it matters: it is a sign of where Chinese wealth is moving. Zhang Yiming is associated with ByteDance, and Liang's debut puts an AI model lab founder on the same list. A rich-list entry is a snapshot of estimated paper wealth, though, not proof of how a company is performing or what its products can do. Read it as a signal of investor and public attention on AI, not as advice about any tool.

What this means for you: students and job seekers can treat it as one more sign that AI skills are in demand, and check actual job listings for specifics.

‘DeepSeek and others’ care about computing power, not Meta's bombshell (Informatization Observer)

An Informatization Observer piece argues that DeepSeek and similar labs are focused on computing power, not on what it calls Meta's bombshell. We only have the headline, so we do not know what the bombshell is or what evidence the author offers. That claim is REPORTED as the author's view; the substance is STILL OPEN.

The framing still points at something real. For model builders, access to chips and the ability to run them efficiently set the pace of progress, which is why a paper like OasisKV above is about memory and not just model quality. For ordinary users, compute limits show up as usage caps, queue times and prices. If the labs you rely on are short of compute, expect rate limits and plan changes before you see better models.

What this means for you: if a tool is essential to your work, read its usage limits now and keep a backup option ready.

Pegasus spyware founder urges sovereign AI (TechRadar)

TechRadar reports that the founder of the company behind Pegasus spyware is telling governments to push for sovereign AI and national security. Sovereign AI means a country building or controlling its own AI systems and infrastructure rather than depending on foreign providers. The headline gives no named speaker, setting or specific policy ask, so those details are STILL OPEN; the call itself is REPORTED.

Why it matters: the speaker sells surveillance tools, so the pitch comes from someone with a commercial interest in the security framing. It also lands next to today's lead story, since both are about who controls powerful software and what it can reach. Read the argument for what it is, a pitch, and weigh it accordingly. Practical takeaway: where your data lives and who can access it is a real choice for businesses. Check where your AI tools process and store your files before you upload client or customer information.

What this means for you: small business owners should ask each AI vendor where data is stored, and keep sensitive files out of tools that cannot answer.

DeepSeek in three headlines, three levels of evidence

Headline one: the ranking. Vico Hao (OFweek) runs “DeepSeek tops the world rankings? Don't rush to pop the champagne.” Which ranking, and by how much, is not in the excerpt we have. Claim and skepticism are both STILL OPEN. OFweek

Headline two: the team. ifeng Tech, with Futu Niuniu also covering, reports sources saying DeepSeek is greatly expanding its team. REPORTED, anonymous insiders, no numbers. ifeng Tech · Futu Niuniu

Headline three: the compute view. An Informatization Observer opinion piece argues that labs like DeepSeek care about computing power, not Meta's headline releases. It is thin on new facts but names the real constraint. REPORTED as opinion. Informatization Observer

Analysis (ours, not the sources'). Hiring is not shipping: a bigger team does not tell you a new model is coming or what license it will carry. Put the three together and you see a lab in growth mode, spending on people and chasing compute. For you, a growing lab usually means a steady stream of releases to evaluate, not a promise of one. Nothing in today's stories says a new DeepSeek model was released.

What to do Tuesday:

• Treat a ranking headline as a prompt to test, not a result.
• When a release lands, read the license before the benchmark table; commercial-use terms are what bite later.
• Check that the weights plus cache fit your memory at your context length.

OasisKV: moving the KV cache past GPU memory

The paper is “OasisKV: Scaling In-Decode KV Cache Beyond HBM with Lookahead Sparse Prefetching” (arXiv 2608.08097, version 2, a revised preprint). arXiv

What the abstract excerpt says (CONFIRMED): LLM serving is increasingly constrained by memory rather than compute, and long-context and long-form reasoning workloads make that worse.

Background (ours): a model generates one token at a time, and each new token looks back at every earlier one. To avoid recomputing, it stores two vectors per earlier token per layer, keys and values. That store is the KV cache. It grows with every token of context and every simultaneous user. HBM (high-bandwidth memory) is the fast memory on data-center GPUs, and it also holds the model weights. In the decode phase, the GPU's math units often wait on memory.

The title's idea, unpacked: sparse means each new token attends to a subset of the cache, not all of it. Prefetching means moving data before it is needed. Lookahead means predicting which pieces the next steps will want. Together: keep the full cache in a larger, slower tier, and while the GPU works on step N, copy in what step N+1 will probably use. If predictions are good, the slow tier hides behind compute. Which slower tier the paper uses is not in our excerpt.

Likely catches (analysis, not the paper's findings): wrong guesses cause stalls; sparse attention can drop information, so quality must be measured; transfers between tiers have a bandwidth ceiling.

Status: the problem statement is CONFIRMED by the abstract. Speedups, models tested, quality loss and code are STILL OPEN; none are in our excerpt, and we do not know of peer review. Novelty is unverifiable from an abstract: offloading and sparse attention are existing ideas as we understand the field, so the claim likely sits in the lookahead combination.

Use today: nothing to download. It does explain why long context is the expensive part of self-hosting, which is the arithmetic in the cost section below.

Lesson, not news: put a coding agent in a smaller box. Illustrative setup; we have not run it, so adjust to your tools. Estimated time: about half an hour once (our estimate).

Who it is for: anyone letting an agent run commands on a machine with real data. Cost: price not stated; check the current terms of whatever agent and container tool you use.

1. Isolate the work. Create a throwaway branch or a separate git worktree so the agent never touches your main work.
2. Hide secrets. No cloud credentials, no SSH keys, no real .env file in the project folder.
3. Contain the commands. Run what the agent executes (builds, tests, installs) in a container with only the project folder mounted. Example: docker run --rm -it -v "$PWD":/work -w /work node:22 bash. Many agents run commands on the host by default. Check whether yours supports a container or remote-execution mode; if not, run the whole agent inside a container or spare machine and allow network access only to your model provider.
4. Install first, then cut the network. Add --network none after dependencies are in place, so install scripts cannot call out.
5. Log everything. If your agent supports hooks, use one to write every shell command to a file.
6. Review. Read the diff, the lockfile, new hidden files and CI config.

Check it worked: plant a canary, a fake file named secrets.env containing the word CANARY. Give the agent a normal task, then search the log for any read or send of it. Absence is a good sign, not proof.

What you give up: convenience. With the network off, some tasks fail, so open one door at a time. Cindy variant: run the same task through Claude Code and Codex in the same box and compare diffs and logs.

Lesson, not news: do the cache math before you buy memory. Using the KV cache idea from the deep dive, here is a calculation you can run on paper. The model below is hypothetical, and all results are estimates.

Formula: 2 (keys and values) × layers × KV heads × head dimension × bytes per value × tokens.

Example: 32 layers, 8 KV heads, head dimension 128, 16-bit values (2 bytes). Per token: 2 × 32 × 8 × 128 × 2 = 131,072 bytes, about 128 KB. At 32,000 tokens of context: about 4 GB per conversation, roughly a feature-length movie file. Ten simultaneous conversations: about 40 GB for cache alone, before the model's weights.

Two levers:

• Store the cache in 8-bit: halves it to about 2 GB per conversation, about 20 GB for ten (estimate; depends on engine support).
• Cap context at what you need: 8,000 tokens uses a quarter of 32,000.

Look for cache-precision and context-length settings in tools such as Ollama, LM Studio or vLLM, which come from different projects and companies. Check each one's docs for exact setting names; support varies. No prices are given in today's stories, so check the current price on any hosted provider's pricing page.

What you give up: possible quality loss from a lower-precision cache, and shorter memory per conversation. Test on your own tasks before switching.

Lesson, not news: a prompt clinic for choosing an open model.

Weak prompt:

What's the best open model for my company?

Why it fails: no hardware, no license needs, no context length, and no way to check the answer, so the model fills the gaps with confident guesses.

Better prompt (illustrative, not tested by us):

You are helping me choose an open-weight model for one specific job. Before recommending anything, ask me any questions you need answered.
My task: [e.g. summarise support tickets in English and Spanish]
My hardware: [GPU model and memory, or 'CPU only' plus RAM]
Context length I need: [tokens]
Concurrent users: [number]
Can data leave my network: [yes/no]
Commercial use: [yes/no]
Give me a table of up to 4 candidates with these columns: model family, size, license name and the clause that matters for commercial use (quote it only from license text I paste below), estimated memory for weights plus KV cache at my context length with the arithmetic shown, and biggest risk.
Label every number either 'from the model card I pasted' or 'my estimate'. If you don't know a fact, write 'unknown - check the model card' instead of guessing.
Finish with a 10-prompt test plan I can run on my own data.
[Paste license text and model cards here]

Why it works: it asks for questions first, pins every claim to a source or a label, forces visible arithmetic, and limits license quotes to text you supplied.

Customize: add a latency target or a language list. Quality check: read the license yourself and redo the memory arithmetic with the formula above.

General lesson, not a new paper: one leaderboard is not a verdict. Today's ‘DeepSeek tops the world rankings?’ headline (OFweek) is a reminder. A benchmark is a fixed set of test questions. A ranking averages scores on those tests, which may not look like your work. Models also vary run to run: ask the same question twice and you can get different answers (that is variance).

Decision this week: before switching models because of a ranking, write 20 real tasks from your own work and run each three times on the current model and the candidate. Count passes. Switch only if the candidate clearly wins on your tasks and its license fits your use. Status: the ranking claim is STILL OPEN; the testing method is general practice.

Three threads run through today. Agent control: as agents gain terminal access and add-ons, defensive tooling and sandboxing become part of the stack. Labs scaling up: reports of hiring and compute focus at DeepSeek point to growth, not a confirmed release. Memory as the limit: papers like OasisKV treat memory, not math, as the serving bottleneck.

Watch next: (1) whether the dev.to author shows what the agent antivirus catches (STILL OPEN, no date); (2) whether any DeepSeek release follows the hiring report (analyst expectation, no date, unconfirmed); (3) whether OasisKV's authors publish speedups and code (STILL OPEN).

Learning with a reason: today it was an estimated halving of a cache bill and a sandbox setup, from sources in China and elsewhere, as of this morning.

I gave my coding agent a sandbox. It immediately asked for network access to find a bigger one.

That's it for today. Tomorrow we're watching whether the agent-security post shows what its antivirus catches, whether a DeepSeek release follows the hiring report, and whether the memory paper shows real numbers. If this made you smarter, share it with one person who would find it useful. Thank you for spending your time with us.

SI Agent News — a podcast newsletter. Made by AI.