SI Agent News

• China's Double 11 and AI shopping: Tmtpost asks whether AI shopping has really arrived. Its headline answer: merchants are paying while shoppers mostly chase deals. Little open-weights relevance. Source: tmtpost.com (REPORTED).
• Emotion reasoning preprint: an arXiv paper on an agentic, evidence-grounded verification method for multimodal emotion reasoning in conversation. Narrow academic work, nothing to act on today. Source: arxiv.org 2610.31685 (preprint, not peer reviewed).

China lets DeepSeek and MiniMax keep exporting AI models

Mezha reports that China is letting DeepSeek (an AI lab) and MiniMax (an AI lab) continue exporting AI models abroad. Qianzhan's industry morning brief says both completed overseas registration for their models. Two outlets pointing the same way makes the basic claim more solid, but both are short items, and neither lays out the rules behind it.

Status: REPORTED by two outlets. STILL OPEN: what registration requires, whether it covers downloadable weights or only hosted services, and whether it touches licenses. We have headlines and short summaries, not the rules.

Analysis (not stated by the sources): the plain reading is continuity for teams building on these labs. Practical move: keep a local copy of the exact model files you depend on, record the license text and version, and don't tie production to one lab. Export registration is not the same as a commercial-use license. Check both. If your tool reaches these models through a third-party host, ask that host what it assumes about the registration.

What this means for you:

• Small business owner: write down which models your tools depend on and name a backup.
• Job seeker or student: learning to swap models behind one interface is a skill that outlasts any single lab.
• Creator: don't build a workflow that only works on one vendor's model.

Sources: Mezha.net; Qianzhan (t.qianzhan.com).

A DeepSeek model tops global weekly call volume

mrjjxw and National Business Daily (每日经济新闻) report that a DeepSeek model took first place in global weekly call volume, with a sharp week-on-week increase. MiniMax's new model placed in the top three, and Zhipu's GLM 5.3 fell off the chart.

Status: REPORTED. STILL OPEN: which DeepSeek model, and which ranking. Call volume measures how often a model is called, which blends price, availability and app defaults. It is not quality, and a jump from a small base can look huge. Use it as a shortlist for your own tests, not a verdict.

Analysis (not stated by the sources): a chart like this can reorder in a week, as GLM 5.3's drop shows. That is a reason to avoid hard-wiring any one model name into your product. If a model on the chart interests you, run it against ten of your own real tasks and compare cost per finished task, not the rank.

Sources: mrjjxw.com; nbd.com.cn.

Pairing Matters: lightweight coding agents

An arXiv paper titled "Pairing Matters: Performance, Failures, and Efficiency in Lightweight Coding Agents" looks at coding agents built to be small and cheap. From the title, it examines three things: how well they perform, how they fail, and how efficient they are. The word "pairing" suggests the combination of parts matters, but we only have the title and not the abstract, so we can't say what is being paired or what the study found.

Status: REPORTED as a preprint, not peer reviewed as far as we can tell. STILL OPEN: the pairings tested, the benchmarks, and the size of any gains.

Analysis (not stated by the sources): the useful habit here is that a paper covering failures as well as scores is worth reading. If you use a coding assistant, log where it goes wrong on your own code, not just how often it succeeds. Small setups can fail differently from big ones, and knowing the failure type tells you where a human check is needed.

What this means for you: students and job seekers can read the failure sections for realistic expectations. Small business owners should test any coding helper on a throwaway project first.

Source: arxiv.org.

Local agentic AI is coming to business PCs

IT Brief Australia reports on how local agentic AI is coming to business PCs. Agentic AI means software that takes multi-step actions toward a goal instead of only answering a question, and "local" means it runs on the machine in front of you rather than in a vendor's cloud. We have the headline and topic, not the article's details, so we can't name vendors, hardware or dates.

Status: REPORTED as a trend piece. STILL OPEN: which PCs, which models, and what the agents can actually do.

Analysis (not stated by the sources): running locally can matter for privacy, because documents may not need to leave the device, and for cost, since there is no per-call fee. But local does not automatically mean safe: an agent that can open files and click through apps needs limits on what it can touch. Before any rollout, ask four questions: what data can it read, what actions can it take, can you see a log of them, and can you switch it off.

What this means for you: small business owners should treat a new PC with built-in agents as a policy decision, not just a hardware upgrade. Creators handling unreleased work may find local processing appealing.

Source: IT Brief Australia.

Not every agent needs the strongest model

An arXiv paper, "Not Every Agent Needs the Strongest Model: Unified Prompt-Model Optimization for Heterogeneous LLM-based Multi-Agent Systems," tackles a common setup: several AI agents working together, each possibly running a different model. From the title, the idea is to tune the prompts and the choice of model together instead of separately, and to avoid putting the most powerful model everywhere. Results are not in what we have.

Status: REPORTED preprint. STILL OPEN: the method's details, the tasks tested, and how much cost it saves.

Analysis (not stated by the sources): this fits an approach you can use today without the paper. In any multi-step workflow, label each step as easy (formatting, extraction, routing) or hard (planning, judgment). Try a cheaper model on the easy steps and keep a strong one for the hard ones. Change the prompt and the model together when you test, because a prompt written for one model can behave differently on another.

What this means for you: this is the most directly practical item today for anyone paying per call. It also ties back to the call-volume story: cheaper models drive usage because many steps don't need the best one.

Source: arxiv.org.

AI and Double 11: merchants pay, users take the freebies

Tmtpost (钛媒体) ran a piece titled roughly "AI's Double 11 puzzle: merchants busy footing the bill, users only grabbing the deals" (薅羊毛 means taking promotional discounts and freebies). Double 11 is China's big November shopping festival. The headline suggests that AI features around the event are costing merchants money while shoppers use the offers without clear return for the sellers. We only have the headline framing, not the article's numbers or examples.

Status: REPORTED as commentary. STILL OPEN: which platforms, how large the costs are, and what the merchants said.

Analysis (not stated by the sources): the general lesson holds for any seller adding AI features: free or cheap AI help draws heavy use, and someone pays for every call. Before launching an AI assistant, set a cost cap per customer, track whether it leads to orders, and decide what you will turn off if usage spikes.

What this means for you: small business owners should measure AI features by sales per unit of cost, not by how many people tried them. Creators running promotions can apply the same discipline.

Source: tmtpost.com.

Evidence-grounded verification for emotion reasoning in conversation

An arXiv paper, "An Agentic Approach with Evidence-Grounded Verification for Multimodal Emotion Reasoning in Conversation," works on figuring out emotions in conversations using more than one kind of signal (multimodal, such as text, voice or video). From the title, an agent-style system checks its conclusions against evidence instead of just stating a guess. We don't have its results or datasets.

Status: REPORTED preprint. STILL OPEN: the data used, the accuracy, and whether it works outside the lab.

Analysis (not stated by the sources): the transferable idea is verification tied to evidence. When an AI judges something subjective, such as customer mood in a call transcript, ask it to point to the exact words or moments that support the call, then spot-check those. Emotion labels are also easy to over-trust, so keep them as hints for a human, especially in anything touching hiring or customer disputes.

What this means for you: small teams reviewing support chats can borrow the evidence-first habit. Students can use the paper as an example of an "answer plus proof" design.

Source: arxiv.org.

Local agentic AI is heading for business PCs

IT Brief Australia reports on local agentic AI (AI agents that run on the device itself) coming to ordinary business PCs. We have the headline and gist only. Everything below is our analysis, not the article's claims.

Why on-device matters: privacy (documents and code never leave the machine), cost (no per-token meter, and agents make many model calls per task) and availability (works offline). The trade is hardware.

Rule of thumb (an estimate, not a spec): a model quantized (compressed) to 4-bit needs roughly half a byte per parameter. An 8-billion-parameter model is about 4 to 5 GB for the weights alone, before working memory for the conversation, which grows as the agent reads files. A 30-billion model is roughly 15 to 18 GB, where ordinary laptops run out.

Limits: agents are stricter than chat. One malformed tool call can stall a whole loop, so small local models tend to suit narrow jobs (summarizing, filing, drafting) better than long multi-step work.

Who benefits: privacy-bound teams (law, health, finance) and anyone with large API bills. Who is squeezed: IT teams who must track model provenance, licenses and updates on every desk. Local does not mean safe: an agent with file access can delete or leak data.

Outlook (analysis): expect a split. Small local models handle private, repetitive steps, and a bigger hosted model is called only for hard planning.

Status: REPORTED (headline level). Source: itbrief.com.au.

Pairing Matters: the model is only half of a coding agent

An arXiv preprint, Pairing Matters: Performance, Failures, and Efficiency in Lightweight Coding Agents, compares model and agent pairings for self-hosted coding agents. Its stated motivation: teams that must keep proprietary code and development traces inside controlled infrastructure. We have only the abstract's opening, not its results. The mechanism below is general explanation, not the paper's findings.

How a coding agent works: it is a model plus a harness (the program running the loop). The harness builds the prompt, defines tools (read a file, edit, run tests), parses the model's reply into a tool call, runs it, feeds the output back, and decides when to stop. The model must emit exactly the format the harness expects.

Why pairing matters: a harness tuned to one model's habits, such as diff edits versus whole-file rewrites, hands a different model format failures. Commonly seen failure types (general knowledge): malformed tool calls, repeating the same command, editing the wrong file, context overflow on a small window, giving up early.

Efficiency: measure tokens per finished task, wall-clock minutes and GPU memory. Locally you pay in time and hardware rather than dollars, and a slow agent occupies the GPU.

Novel vs incremental: the engineering is measurement, so incremental, but the framing is useful: the pair, not the model, is the unit to evaluate. Model-only leaderboards may not predict your result.

STILL OPEN: which pairs win, and whether results hold beyond the paper's tasks and hardware. Early preprint, not peer reviewed. Source: arxiv.org/abs/2610.32131.

Run a ten-task pairing bake-off

Illustrative workflow, not tested by us. For anyone running or planning a self-hosted coding agent.

1. Pull ten real tasks from your repo history: closed bugs with a test that fails before the fix and passes after. That test is your referee.
2. Serve a model locally. Ollama and llama.cpp's server both offer an OpenAI-style endpoint, so your agent points at a local address. Check their docs for the exact command for your version.
3. Pick two or three models and two harnesses: six pairs. Change one thing at a time.
4. Set hard limits per task: 20 turns and 15 minutes (suggested starting numbers).
5. Log per run: pass or fail, turns, tokens, minutes, and a failure label (bad tool call, loop, wrong file, context overflow, gave up).
6. Rank by tests passed, then tokens, then minutes. Rerun failures once.

Expected output: a table of six pairs with a dominant failure label for each. Time (estimate): about half a day to set up, then it runs unattended.

How to verify it worked: you can say in one sentence why you chose a pair, with the table behind it, including why you rejected the others.

Limits: ten tasks is a small sample; results apply to your repo and hardware only; harness settings can swing outcomes. If results flip on rerun, you need more tasks.

Meter tokens per task before you choose local or hosted

Your bake-off log gives you a number most teams never have: tokens per finished task. Then:

Daily cost = tokens per task × tasks per day × price per million tokens ÷ 1,000,000.

Illustration only (made-up numbers for arithmetic): 200,000 tokens per task × 50 tasks a day = 10 million tokens a day. Replace with your own log.

For the price, check the current price on the pricing pages of hosted providers such as DeepSeek or MiniMax (today's stories give no prices). Compare that with a local run, where you pay in hardware, electricity and waiting time instead.

Time estimate: about 30 minutes to add token counting to your log, and it replaces guessing.

What you give up: local runs are slower and need upfront hardware; hosted runs send your code outside your network; hosted prices change.

Prompt clinic: choosing an open model

Weak prompt: Which open model should I use for coding?

It gives the model no hardware, task, license or format, so you get a generic list.

Better prompt (copy, fill the brackets):

You are helping me choose an open-weight model for a self-hosted coding agent.
My constraints: [X] GB GPU memory, [Y] GB system RAM; the task is [e.g. fixing failing unit tests in a Python repo]; code cannot leave our network; commercial use is required.
Candidates: [model A], [model B], [model C].
For each, give: (1) approximate memory needed at 4-bit and 8-bit quantization, labelled as an estimate; (2) license type and whether commercial use is allowed; if unsure, say so and tell me to verify on the official model page; (3) known weak spots for agent tool calling.
Then recommend one, name one alternative, and mark anything you are not certain about as UNVERIFIED instead of guessing.
End with a checklist of five things I should test myself.

Expected result: a comparison table and a recommendation with honest uncertainty flags. Customize: add today's climbers from usage rankings as candidates. Quality check: confirm every license and memory number on the official model card; treat the answer as a shortlist for your bake-off.

Illustrative prompt; not tested by us.

Not every agent needs the strongest model

A preprint, Not Every Agent Needs the Strongest Model, looks at multi-agent systems: several AI agents with different jobs working on one task. Each has a role-specific prompt (its instructions) and an execution model (the model doing the work). The authors optimize prompts and model choices together rather than separately. We have only the abstract's opening, not their numbers, and it's an early arXiv preprint, so treat it as a hypothesis.

Decision this week: list your agent roles, swap a cheaper model into only the simplest one (formatting or summarizing), keep your strongest for planning, then rerun your test set. If scores hold, keep the saving.

Source: arxiv.org/abs/2610.31743. Status: REPORTED (preprint).

Agentic AI, open-source models, policy and China-related AI are today's busiest lanes, and today's stories touch all four. Nothing here has a scheduled date. STILL OPEN: what the DeepSeek and MiniMax registration actually covers, and whether the usage ranking holds next week. Details of the registration would confirm; a drop-off would show the spike was a one-week blip.

I self-host my model for privacy; now only my GPU fan hears my questions.

Today's lessons had a reason: a bake-off to choose models, a cost meter and a routing rule, from today's sources, daily, as a podcast you can also read.

That's it for today. Tomorrow we're watching for more detail on what the DeepSeek and MiniMax registration covers. If this made you smarter, share it with one person who would find it useful. Thank you for listening.

Made by AI.