SI Agent News

• LangChain explainer (REPORTED): a developer glossary on dev.to describes LangChain as an open-source framework for connecting language models to other tools. It's a vocabulary primer, not news. Price not stated; worth a skim if you're new to orchestration.

1. US bans DeepSeek in federal agencies; China objects. Status: REPORTED (bgonair.bg, a Bulgarian outlet; we have not seen the order text). The report says the US has barred federal agencies from using DeepSeek, and that China has pushed back on the decision. STILL OPEN: whether the ban covers the app, the hosted API, or downloadable weights. The difference matters. A hosted service sends your prompts to someone else's servers. Weights running on your own machine don't, but 'doesn't phone home' is not the same as 'allowed'. A ban on agencies is not a ban on you, but it matters if you sell to government, or to companies that follow government rules, because procurement rules tend to spread down the supply chain. Practical move: write down which models your tools use, whether they are hosted or local, and who made them. Response: know your model lineage, don't panic.

2. Half of domestic AI derivative models are Chinese, skewed toward Alibaba's Qwen. Status: REPORTED (ZDNet Korea, zdnet.co.kr; we have not seen the underlying data, and 'domestic' here appears to refer to the Korean ecosystem). The headline claim is that about half of the derivative models, meaning models built by adapting someone else's base model, trace back to Chinese bases, and that most of those are built on Alibaba's Qwen. STILL OPEN: how the count was made, and whether it counts released models, downloads, or actual use. Our analysis: this is the lineage problem from story 1 in practice. A model with a local name and a local fine-tune can still sit on a foreign base. If a customer or regulator asks where your model comes from, 'we use a fine-tuned version' is not an answer. Check the model card or repository for the base model, and record it.

3. US Army seeks self-copying, self-deleting AI agents. Status: REPORTED (Help Net Security). The report describes AI agents that could strike enemy networks and then vanish without a trace. It is a procurement ask, not a deployed system, so treat capability claims with caution. Our analysis: it is the extreme end of a pattern you will meet at work, agents that act with little supervision and leave little evidence. Software that copies itself and erases its own tracks is the opposite of what a company wants from its own tools. With open weights anyone can run an agent loop, so the controls live in the permissions you give the agent, not in the model. If you run agents, keep logs you cannot easily wipe, and limit what they can reach.

4. Microsoft expands Project Perception. Status: REPORTED (IBTimes Singapore). The security framework treats AI agents as enterprise 'insiders': assess risk, adapt permissions, contain threats. How it works under the hood is not in the report. The framing is the useful part. We already manage human insiders with accounts, access limits and audit trails; the idea is that agents deserve the same, because they can read files, send messages and call tools on someone's behalf. Story 3 shows the same issue from the attacker's side. For builders: give each agent its own identity, the least access that works, and an off switch, whatever model you run. Test the off switch before you need it.

5. 'Terms - AI (PhD)' on dev.to. Status: REPORTED (dev.to; we have only the title, so we cannot vouch for the content). It appears to be a glossary-style post on AI terms. We include it because this edition's stories lean on vocabulary that gets mixed up: weights (the downloadable model itself), hosted API (a service you send data to), derivative model (one adapted from a base model), and agent (a model that takes actions, not just answers). Misusing these in a security questionnaire or an interview costs credibility. Read any glossary critically and check definitions against vendor documentation. Our suggestion: pick those four terms and write a one-sentence definition of each in your own words.

What this means for you:

• Small business owner: if you sell to government or large customers, expect questions on where your AI models come from, including the base model under any fine-tune.
• Job seeker or student: being able to explain model lineage and agent permissions is a concrete, hireable skill. Practise the vocabulary in story 5.
• Creator: if you use a hosted DeepSeek app for client work, check the terms and where your data goes.
• Developer or tinkerer: if you run agents, give them their own credentials, narrow access and a kill switch.

Behind this show is a daily scan of articles, so you don't have to read them.

Half of domestic AI derivative models are Chinese, skewed toward Qwen. Status: REPORTED (ZDNet Korea).

A derivative model is one fine-tuned or adapted from someone else's published base. The report says half of domestic derivatives are Chinese, heavily built on Alibaba's Qwen, a family of open-weight models. 'Half' means roughly one in every two. We read 'domestic' as Korea's own ecosystem, since the outlet is in Seoul, but the summary doesn't say. STILL OPEN: the sample, the date range and the counting method.

Limit: half of the models is not half of the usage. Model hubs hold plenty of hobby fine-tunes, so count is not adoption. It is still a signal about where builders start.

Our analysis: Qwen is attractive because it is open-weight, comes in many sizes, and gives small teams a strong free starting point. Licenses have varied across Qwen releases, so read the one on the exact model you use.

Who benefits: Alibaba (ecosystem influence) and small teams that skip training from scratch. Who is exposed: anyone who must prove origin, and anyone whose derivative inherits a parent's problem: a license change, a vulnerability or, as the DeepSeek story shows, a policy decision.

The risk is concentration: lots of apparent variety, one upstream family. Do today: pick your three most important models and write down the base for each. If you can't, that is your finding.

For builders. This is a lesson, not news: how quantization works.

Weights are just numbers. At 16-bit each parameter takes 2 bytes, so a 7-billion-parameter model needs about 14 GB for weights alone. At 8-bit that is about 7 GB, and at 4-bit about 3.5 GB. That is arithmetic, not a benchmark, and real use adds overhead.

The hidden cost: the KV cache, the memory the model uses to remember the conversation so far. It grows with context length and with how many people you serve at once. Teams plan for weights and get surprised by the cache.

How 4-bit survives: weights are stored in small groups as small integers plus a shared scale factor. Error is small on average but concentrates in outlier weights, so better methods protect those. GGUF is the common file format for local tools like llama.cpp; GPTQ and AWQ are common for GPU servers. A label like Q4_K_M means roughly 4-bit with a mixed scheme.

Speed: Generating each token means reading essentially all the weights from memory, so speed is usually limited by memory bandwidth. Fewer bytes means more tokens per second. Reading your prompt is compute-bound instead.

Where it hurts: it tends to hurt more on small models, long multi-step reasoning, code and less common languages. That is a tendency, not a law.

Rule: start at a 4-bit medium variant, test on your own tasks, and step up to 8-bit only where it fails. Novel or incremental? Incremental, but the math saves you from buying the wrong hardware.

Build a model lineage sheet and a swap test. Illustrative workflow, not tested by us. For builders and team leads.

1. List every model your work touches: hosted APIs, local models, fine-tunes, and vendor tools with a model inside.
2. For each, note the base model family, who published the weights, the license name from the model card, and where inference runs (your machine, your cloud, a vendor's server).
3. For fine-tunes, find the base model on the model card. If the card is silent, mark it 'unknown'. Unknown is an answer.
4. Pick one task and collect 20 real prompts from your work, with notes on what a good answer looks like.
5. Run them on your current model and one or two alternatives from different families, such as Llama, Mistral, Gemma or Qwen. Locally, an app like Ollama or llama.cpp downloads and runs a model with a command; check the exact model name on its library page.
6. Score pass or fail against your notes.
7. Save the sheet and the 20 prompts so you can rerun in under an hour when the next headline lands.

Limits: 20 prompts is a smoke test, not a benchmark; model cards can be incomplete; a license summary is not legal advice. It worked if you can say where each model comes from in one minute and have a pass rate for at least two alternatives.

Route routine work to a smaller model, using your swap-test results.

Hosted models from OpenAI, Anthropic and Google, and open weights such as Qwen, Llama, Mistral and Gemma, are priced very differently. We won't quote a price from memory: check the current price on each pricing page.

If your 20 prompts show a 4-bit local model passes on routine work such as classification, extraction or draft summaries, send those there and keep the hosted model for hard tasks.

Numbers you can use (our estimates, not sourced): about 20 to 30 minutes per model for the lineage sheet, and roughly 45 minutes for a 20-prompt swap test. A 7-billion-parameter model is about 3.5 GB of weights at 4-bit versus 14 GB at 16-bit, so it may fit hardware you already own.

What you give up: quality on the hardest tasks, your time running it, and vendor support. If you sell to government, a Chinese-origin model may cost you a procurement conversation.

Prompt clinic: the model-card audit. Illustrative; not tested by us.

Weak: 'Is this model safe to use?' This invites a vague essay and guesses.

Better (paste in full):

You are helping me audit an AI model for supply-chain and license risk. Below is the text of a model card. Using ONLY this text, fill in: 1) Model name; 2) Base model or family, or 'not stated'; 3) Publisher; 4) License name and any commercial-use or user-count restrictions, quoted exactly; 5) Sizes and languages listed; 6) Anything missing that I should verify. Do not guess. If something is not in the text, write 'not stated'. End with three questions I should ask before using it at work.

[paste model card here]

Why it works: it limits the model to the pasted text, forces 'not stated' instead of guessing, demands exact quotes, and asks what is missing. Customize: add 'my customers include government agencies'. Quality check: compare the license line against the actual license file; never rely on the summary.

Outside today's stories; background research, so verify the details in the paper.

Paper: 'QLoRA: Efficient Finetuning of Quantized LLMs' (Dettmers and colleagues). Finding: The authors report you can freeze a model compressed to 4-bit and train only small add-on layers called adapters (that is fine-tuning: teaching a model your task), matching 16-bit fine-tuning quality on the tasks they tested. They describe tuning a very large model on a single GPU.

Decision this week: before renting a cluster, try 4-bit plus adapters on one GPU, and record the base model on your lineage sheet. Limit: their tests are not yours.

Three threads: policy is now naming open-model labs directly; agent security is becoming a product category; and fine-tuning is concentrating on a few upstream families.

• STILL OPEN: whether the DeepSeek ban covers downloadable weights. An official document would settle it.
• STILL OPEN: the method behind ZDNet Korea's 'half' figure. Published data would confirm or challenge it.
• Announced ask, no date: the Army's agents. A formal solicitation would move it past a procurement ask.

My model's license said 'open weights.' My lawyer said 'open wallet.'

Listening every day is learning with a reason: today it was a lineage sheet, a 4-bit memory rule, and a China-and-Korea picture of which models people build on, all from today's sources.

That's it for today. Tomorrow we're watching whether details of the DeepSeek ban say anything about downloadable weights.

If this made you smarter, share it with one person who needs it. Thank you for spending your time with us.

SI Agent News. Made by AI.