SI Agent News

• Anthropic published a post titled 'Investing in the public benefit' (anthropic.com). We only have the headline, so we are not characterizing the contents. STILL OPEN.
• China's Financial Street Forum opens today, with the heads of four financial regulators due to speak, and leading tech stocks have seen heavy buybacks (Cailian Press, cls.cn). Market context only; the AI link is indirect. REPORTED.
• deepseek-harness, an open-source plugin-style project, is described in a dev.to post as having 253,062 stars. We could not verify that number and the post reads templated. Check the repository yourself before installing. STILL OPEN.

DeepSeek's V4.1-Flash, a language model, claimed lower HBM needs, and Korean chip stocks fell.

HBM is high-bandwidth memory, the fast memory stacked beside AI chips. According to Crypto Briefing, South Korean stocks slid, Samsung and SK Hynix (the two big memory chipmakers) were hit, and the KOSPI (South Korea's main stock index) was dragged lower. Yahoo Finance Hong Kong adds that Japanese shares also fell and that AI-financing worries were weighing on chip stocks alongside the new DeepSeek version.

Status: the market move is REPORTED by two outlets. The lower-memory claim is DeepSeek's own, and neither report gives a number or a baseline, so it is STILL OPEN until someone measures it independently. Today's coverage does not state the license or whether the weights are fully downloadable. Check the model card.

Analysis: memory is the main limit on what you can serve, so a credible cut changes which model fits on which card. But cheaper per request often means more requests, not fewer chips. Nobody knows yet. A one-day share move on an announcement shows how nervous investors are. It does not show measured demand for memory.

What this means for you:

• Small business owner: hold off on buying hardware until independent tests of V4.1-Flash land.
• Job seeker or student: learning to size and run models locally is a skill you can show in a portfolio.
• Creator: lighter models could eventually mean running tools on your own laptop. That is a maybe, not a promise.

DeepSeek upgrades its programming model and opens a paid search API.

A report on mrjjxw.com says DeepSeek upgraded its programming model and opened a search API priced at 5 yuan per thousand calls. That works out to 0.005 yuan per call, or about 500 yuan for 100,000 calls. The same headline says Anthropic plans to raise 60 billion dollars. The headline is cut off before saying what for, so we do not know the purpose.

Status: REPORTED by a single outlet. The price comes from that report, not from a DeepSeek page we have seen. We also have no name, benchmark or license for the upgraded coding model, and no confirmation of the Anthropic figure.

Analysis: a priced search API lets an app pull in fresh web results without building its own crawler. Pricing per thousand calls makes the cost easy to forecast, but you still need to know what one call returns and what limits apply.

What this means for you: if you build tools, check DeepSeek's own pricing page before budgeting. Treat the Anthropic number as a rumor until the company or a second outlet confirms it.

Anthropic publishes 'Investing in the public benefit.'

Anthropic has put out a post titled 'Investing in the public benefit.' We have only the title and the source, not the text. So we cannot say what it commits to, how much money is involved, or whether it connects to the 60 billion dollar fundraising report above. Do not assume a link.

Status: the post exists, per Anthropic. Its contents are UNREAD here, so nothing in it is verified by us.

Analysis: when a company writes about public benefit, the useful questions are concrete. Does it name amounts and dates? Does it describe something binding or only intended? Who decides, and who can check? Vague language costs the company nothing. Specific, checkable commitments are the ones worth tracking.

What this means for you: if you rely on Anthropic's tools for work or study, read the post yourself. Look for changes to access, pricing or terms rather than general statements of purpose.

Stanza generates stateful mock APIs from a plain-language prompt.

A dev.to post introduces Stanza, a tool that builds instant, stateful mock APIs from natural language prompts. A mock API is a fake stand-in for a real online service, used so you can build and test without touching the real thing. Stateful means it remembers data between calls. If you create a record, a later request can find it, as a real service would.

Status: REPORTED by the author on dev.to, a developer blogging site. We have not run it, and the post's claims are unverified here.

Analysis: stateful mocks are useful because many bugs only show up across several calls, such as create, then update, then delete. The risk is that a mock generated from a prompt may behave differently from the real service. It can silently invent rules the real one does not have.

What this means for you: for a student or job seeker, a mock API lets you build a portfolio app with no accounts or keys. Compare the mock's responses with the real service's documentation before trusting it, especially for anything involving money or personal data.

Financial Street Forum opens, with four financial regulators set to speak.

Cailian Press's morning news says the Financial Street Forum opens today and that the heads of four financial regulatory departments will speak. The headline also says tech leading stocks are doing something, but the text is cut off, so we cannot say whether they rose, fell or moved sideways. We do not have what the officials said.

Status: the forum and the speaker lineup are REPORTED by one outlet. Any policy content is UNKNOWN until the speeches are published.

Analysis: regulator speeches at events like this can shift market mood, because they signal priorities for lending, markets and technology. It is a separate item from the Korean chip selloff, and nothing here links the two.

What this means for you: if you follow Chinese tech stocks or invest in AI-linked funds, read the full transcripts or official summaries instead of headlines. A short quote can mislead, and the first price reaction often reverses.

A dev.to writer tests deepseek-harness.

A dev.to post titled 'I Tried deepseek-harness — Here's What You Need to Know' is a hands-on write-up of something called deepseek-harness. We have only the title. We do not know what it does, who maintains it, or whether DeepSeek itself made it.

Status: a first-person review, so it is one person's experience and UNVERIFIED by us. The name alone does not show any official tie to DeepSeek.

Analysis: small tools that wrap or steer a model can save time, but they also sit between you and your API key. A name that borrows a company's brand is not proof of origin.

What this means for you: before installing any such tool, check four things. Look at who publishes it, which license it uses, what permissions or keys it asks for, and when it was last updated. Try it in a sandbox first, and never give it a key tied to a card or a production account. Then compare the review's claims with what the tool does on your own test task.

DeepSeek puts a price on search, and the same roundup carries Anthropic's $60 billion chip plan.

According to a Chinese morning roundup on mrjjxw.com (dated October 19, China time, which runs ahead of US dates), DeepSeek upgraded its programming model and opened a search API at 5 yuan per 1,000 calls. That is 0.005 yuan a call, so 100,000 searches would cost 500 yuan.

Why search matters: a downloaded model only knows what it learned before training stopped. To answer anything current, an agent has to look things up. A flat per-call price makes that cost predictable instead of buried in token counts. Builders running agents on open models benefit. Search providers priced above it may feel pressure, but that is our inference, not reported.

The gap: the roundup says the coding model was upgraded, but gives no version name, benchmark scores, price or license. The upgrade and the search price are REPORTED from a single outlet. The specifics are STILL OPEN. Do not switch your coding assistant on a headline.

The financing number: the same roundup reports that Anthropic plans to raise $60 billion to buy AI chips, offering Blackstone (an investment firm) a yield above 10%. REPORTED, and a plan, not a closed deal.

Analysis: one side is raising tens of billions for hardware, the other is claiming it needs less memory and selling search by the thousand calls. Those are two bets on where AI costs end up, but they are not necessarily opposed, since cheaper models can raise demand.

Cost and worth: search at 5 yuan per 1,000 calls; coding model price not stated. Worth testing if you build agents that need fresh web information; skip it if your tasks never need current facts.

For builders: the memory math behind 'needs less HBM'. This is a lesson in general engineering, not news. We have not seen V4.1-Flash's architecture, so none of this describes how DeepSeek did it.

Three parts of memory. Weights, KV cache, overhead.

• Weights = parameters × bytes per parameter. At 16-bit that is 2 bytes each, so a 70-billion-parameter model is about 140 GB. At 8-bit, about 70 GB. At 4-bit, about 35 GB. These are estimates before overhead.
• KV cache stores attention data (keys and values) for each token the model has read, so it does not recompute them. It grows with context length and with the number of simultaneous users, and on long documents or busy servers it can rival the weights.

Three meanings of 'lower memory'. Smaller weights (quantization); a smaller cache (designs that share or compress attention data); or fewer chips for the same traffic. A mixture-of-experts model, which activates only some parameters per token, saves compute but still has to store all its weights, a common confusion.

Limits. Quantization trades some quality, and how much varies by model and task. Test on your own prompts.

Local rule of thumb (my estimate): at 4-bit, about 0.5 GB per billion parameters, plus about 20% headroom, plus cache for your longest prompt. A 30B model: 15 GB of weights, about 18 GB with headroom, which fits a 24 GB card. A 70B model: 35 GB, about 42 GB with headroom, which does not.

Why HBM matters: it is the memory on data-center GPUs; consumer machines use VRAM or unified memory. When a vendor says 'less memory,' ask which of the three they mean.

For builders: test an agent against a stateful mock API before it touches anything real.

Tool: Stanza, an open-source developer tool that turns plain-language descriptions into stateful mock APIs (one that remembers earlier calls, so 'create' then 'fetch' returns what you created). It comes from a Hacktoberfest submission on dev.to, and the author calls it production-ready. We have not tested it. Cost: open source; license and model costs not stated, so check its repository. Worth it for: developers building agents or front ends before the backend exists.

Steps (illustrative workflow):

1. Describe the API in plain words: resources, fields, rules, errors.
2. Generate the mock following the project's README. Do not guess install commands.
3. Write a state-sequence test: create an order, pay it, cancel it, fetch it. Expect status 'cancelled'.
4. Point your app or local open-model agent at the mock's base URL.
5. Inject failures: a rate-limit error and a server error. Watch how your agent retries.
6. Replay the sequence and compare against your expected results.

Orders API. An order has id, item, quantity, status. Creating an order sets status to 'pending'. Paying a pending order sets 'paid'. Cancelling a paid order sets 'cancelled' and returns 200. Cancelling an already-cancelled order returns 409. Every 10th request returns 429.

Limits: a mock is only as good as your description, and passing against it does not prove production behavior. Verify: the final fetch shows 'cancelled', the 409 and 429 cases occur, and your agent stops cleanly instead of looping.

Cache your agent's repeat searches. The one real price in today's set: DeepSeek's search API at 5 yuan per 1,000 calls (REPORTED).

Illustrative math (my assumptions): 20,000 searches a month × 5 yuan per 1,000 = 100 yuan. If 40% of queries are repeats, which is an assumption and not a measured figure, so check your logs, a cache cuts the bill to 60 yuan. Savings: 40 yuan, roughly the cost of 8,000 searches.

How: normalize each query (lowercase, trim spaces), store the result with an expiry, and check the cache first. Roughly 24 hours suits stable topics; use much shorter for news.

Options to compare: DeepSeek's search API, Brave Search API, Tavily, Google's Programmable Search. Check the current price and caching terms on each pricing page; we are not stating their prices.

What you give up: stale results, and some providers restrict storing their results, so read the terms first.

Prompt clinic: evaluating a new open-weight model release.

Weak prompt: Is the new DeepSeek model better than what I'm using?

Why it fails: it names no task, no hardware, no budget, and no definition of 'better', so the model guesses or flatters the release.

Better prompt (copy-ready):

I am deciding whether to switch from [CURRENT MODEL] to [NEW MODEL] for [TASK, e.g. code review of Python files]. My hardware: [GPU/RAM or cloud setup]. My budget: [AMOUNT]. Context length I need: [TOKENS].

Using only the facts I paste below, do this:
1. List each claim the vendor makes and label it: measured by a third party, measured by the vendor, or unmeasured.
2. Estimate memory needed at 16-bit and 4-bit using parameters x bytes, and show the arithmetic.
3. List what I still need to find out (license, context length, quality loss from quantization).
4. Write a 20-prompt test plan from my real work, with a pass/fail rule for each.
Do not invent numbers. If a fact is missing, say 'not provided'.

FACTS: [PASTE MODEL CARD AND ANNOUNCEMENT]

Why it works: it forces a claim-by-claim status, exposes the math, and bans invention.

Customize: swap in your task and hardware. Quality check: redo the memory arithmetic by hand for one model.

No new paper in today's set, so this is a lesson, not a finding. Today's lesson: treat a vendor's efficiency claim as a hypothesis. A benchmark is a standard test set with scored answers; peak memory is the most RAM the model uses at any moment; tokens per second is how fast it writes. A vendor's own numbers are not independent proof.

Decision this week: before switching models, run 20 prompts from your real work on both, and record peak memory, tokens per second, and a pass/fail score on quality. About a morning's work (my estimate). If the new model wins on all three, switch. If it wins on memory but loses quality, keep the old one for important tasks.

Open models are now moving markets: one release with a lower-memory claim hit chipmaker shares the same day. Capital intensity is the other side of the story, with Anthropic's reported $60 billion chip plan.

What to watch:

• Confirmed and scheduled: China's Financial Street Forum opens today, with four regulators due to speak (Cailian Press).
• Announced plan, no firm date: Anthropic's chip financing (REPORTED). Confirmation would be a closing announcement.
• STILL OPEN: V4.1-Flash's memory claim. Independent peak-memory measurements would confirm it; none appearing would leave it unproven.

DeepSeek says its model needs less memory — and somehow the stock market forgot everything first.

That's it for today. Tomorrow we're watching for independent tests of V4.1-Flash's memory claim, what regulators said at the Financial Street Forum, and any word on Anthropic's financing.

If this made you smarter, share it with one person who needs it. SI Agent News is independent and self-funded, with no investors and no ads., and every share and every bit of support keeps it going. Thank you.

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