Anthropic Cuts Distillation Extraction by 90–99%, EU Moves on Vietnam Minerals, and Builders Debate Self-Hosting vs. SaaS

Key Takeaways

1. Anthropic says new defenses cut extracted reasoning data by roughly 90–99% and that it banned tens of thousands of accounts tied to distillation attacks, naming Zhipu and Moonshot again.
2. The EU is seeking closer ties with Vietnam to diversify critical mineral supply away from China, per internal documents.
3. China's September trade data showed exports up 36.3% year on year and a record $133.5bn surplus, while PPI rose 5.0% and CPI rose just 0.9%.
4. The IMF raised Taiwan's 2026 growth forecast to 10.7% from 5.2% in April and cut the mainland forecast to 4.5%, citing AI and advanced chip demand.
5. Builders are split on self-hosting: some report large SaaS savings from in-app AI translation and self-built stacks, while others warn that email deliverability and maintenance costs make DIY a trap.
6. Agentic workflows are moving from manual prompting to scheduled routines, with users describing agents running support triage, bug routing, and content tasks in the background.
7. A recurring theme across posts: AI is compressing the cost of building software, but the durable edge is shifting toward distribution, context, and human judgment.

1. AI Models and Infrastructure

1. Anthropic said in an October 12 follow-up post that it has banned tens of thousands of accounts linked to distillation attacks and that its initial estimate shows new defenses reduced extracted reasoning data on its platform by about 90% to 99%. The company said it rejected trillions of tokens of paid traffic, but acknowledged it does not know how long the defenses will hold, that some legitimate requests may be falsely blocked, and that those requests may still be charged. It again named Zhipu and Moonshot — via 1
2. On the infrastructure side, one builder said Anthropic has released strong models and that most people around them now use Opus as their main model, while predicting Grok will catch up and take second place; they described the current ChatGPT as “reset” and “slow.” Separately, a post claimed Anthropic is “beyond” git because it cannot handle 25 million CI tasks per day, framed as evidence that agentic engineering is a quantum leap — via 1 2
3. OpenCompute was shared as a self-hosted orchestrator that runs agent sandboxes on a user-chosen runtime, with the stated goal of letting autonomous software development happen without handing code and context to someone else's cloud, and scaling remote agent clusters inside a VPC — via 1

2. Trade, Macro, and Geopolitics

1. Internal EU documents show the bloc is seeking to strengthen ties with Vietnam to secure critical mineral supply and reduce dependence on China. Vietnam holds large reserves of tungsten, rare earths, and other minerals key to the defense industry, and the EU wants to diversify sources after China's export restrictions last year disrupted European supply chains — via 1
2. China's September trade data showed exports up 36.3% year on year in dollar terms, far above the 25.7% market expectation, with a record monthly trade surplus of $133.5bn. Chips, automatic data processing equipment, and parts accounted for 17.9% of exports and contributed 13 percentage points to export growth; exports to the US rose 39.6% and to the EU 25.3%, while imports from the EU grew just 0.7%. In the same month, PPI rose 5.0% year on year, the largest gain since June 2022, while CPI rose only 0.9%, below the 1.0% expectation, with gasoline prices up 16.6% year on year contributing about 0.47 percentage points. Industrial producer purchase prices rose 6.9% year on year; Reuters noted that rising raw material costs are pressuring manufacturers, but weak domestic demand limits their ability to pass costs to consumers — via 1 2
3. The IMF raised Taiwan's 2026 economic growth forecast to 10.7% from 5.2% in April, and cut the mainland forecast to 4.5%. Taiwan's growth is mainly driven by AI and advanced chip demand, with Taiwan and South Korea together accounting for 62% of net trade in AI-related computing hardware. Chinese Foreign Ministry spokesperson Mao Ning said the fundamentals of China's economy remain sound and that the full-year growth figures are forecasts — via 1
4. The China–North Korea new Yalu River bridge is set to officially open on October 15, after sitting idle for about 12 years since its completion in October 2014. China fully funded the construction; the North Korean side's connecting roads and customs facilities were not completed in time, delaying opening. Xinhua reported in 2014 that total investment was 2.22 billion yuan — via 1

3. Builders, Agents, and the Self-Hosting Debate

1. Tony Dinh said he replaced his i18n translation SaaS with an in-app translation tool, saving $124 per month. His app has 2,705 translation keys across 16 locales, and using GPT-Luna 6 costs about $0.75; he said the translations are more accurate and more context-aware. On the other side of the ledger, Jack Ellis warned against building your own email delivery service: emails landing in spam mean lost revenue and opportunities, deliverability is extremely hard, and it is not something to solve with vibe coding. He also said one of the biggest mistakes of the past year was “building your own software,” arguing that even a decent clone carries maintenance costs heavy enough that he would rather pay — via 1 2 3
2. Arvid Kahl used a prompt telling Claude to implement a checklist item by item to satirize the idea of vibe-coding a beehiiv replacement over a weekend. The quoted post listed a long feature set including 99.3% email deliverability, sending over 3 million emails per minute, three-layer suppression lists, DKIM/SPF/DMARC, REST and sending APIs, an MCP server, automation, payments, community, podcasts, and an ad network, and noted that a self-built version still requires assembling SendGrid/SES, Stripe, Clerk, Cloudflare, Litmus, and many other third-party tools. Kahl argued the key is not whether a usable replacement can actually be built — he called it “too hard” — but that there is now a strong desire to try, which is enough for capable entrepreneurs. A quoted reply said they are cutting many SaaS tools with a self-built stack and that beehiiv was on the list, but even without needing 90% of the listed features, they could not reliably send 65,000 emails at a lower price — via 1 2
3. Jon Yongfook said the switching cost for adopting new technology is now nearly zero, so after almost a decade on Bulma CSS he moved to Tailwind, joking he might try React in 20 years. He also said he was building a SaaS priced at $399 per month, but his agent vibe-coded a replacement for $9 per month, calling the environment brutal. He said he encounters someone every week saying they will weekend-vibe-code a linktree replacement, and listed the actual scope: input, validation, saving, reading, display, editing, deletion, confirmation/cancel deletion, link CRUD, naming, sorting, moving, copy-paste, validation and correction, search and hidden recovery, loading states, link lifecycle, and inactive-user management. He argued that all-in-one platforms that charge too much risk being unbundled by vibe coding, with users rebuilding only the few features they need — an existing unbundling risk that AI makes easier — via 1 2 3 4
4. Tibo described AI agents as feeling like 50 employees standing in the office staring at you, creating a sense of waste and leading to “fried brain” from all-day manual prompting. His response was to stop manual prompting, move workflows to Squad, and build independently running “ROUTINES”: hourly support ticket triage, routing bugs to dev agents, adjusting based on data, replying to outreach in his own voice, generating three tool ideas daily from the X timeline, running AI benchmarks overnight, producing daily videos on schedule, finding leads for outreach, and replying to close deals — all run in a loop led by a lead agent. He also announced his founder community is free but requires proof of $1k/month revenue, with benefits including a Slack of 250+ verified bootstrap founders, channels segmented by revenue stage ($1k+, $10k+, $100k+), and weekly 1:1s with founders at the same revenue level; deeper calls are $99/month with AMAs, member problem deep-dives, and guest sessions from @marclou, @illyism, @wickedguro, and @robj3d3 — via 1 2
5. Ben Tossell shared a case of using @FactoryAI's coding agent to scale design workflows: batch-rearranging design assets for global billboard campaigns in New York, Tokyo, London, and elsewhere, handling grouping, cropping, and unexpectedly placed panels. Formatting all panels to spec previously took hours to days of non-creative work. In a podcast conversation, he described himself as “non-technical” and afraid of the terminal, but said he has shipped code with @FactoryAI; he keeps agent context in folders, does not use Claude Code, and said AI makes him want to spend more time on the computer. The episode also covered assigning different tasks to Pi, Factory, and Codex, preferring to guide agents while working, using a shared vocabulary to reduce communication drift with agents, making personal context portable across agents and models, reducing agents' memory of him, and the view that calling AI adoption a “skills problem” lets AI labs off the hook — via 1 2
6. Daniel Vassallo said this was his first flight in the agentic AI era and he was amazed to use free in-flight WiFi and check, via the Claude app on his phone, a PR being handled by a Claude code instance. He praised Linus Torvalds' pragmatism, saying Torvalds treats curiosity and willingness to learn new tools as useful qualities; the quoted post said Torvalds called rules banning AI use “extremely stupid” and “idiotic,” and that the Linux kernel will not make stupid rules saying you cannot use a certain tool to do something. Vassallo also said some people still do not see agentic engineering as a quantum leap — via 1 2 3

4. Distribution, Content, and Market Signals

1. Tibo said Google is becoming an important source of organic traffic and real revenue growth for him. His approach includes using Outrank to build backlinks and generate content, then using Squad to expand and optimize content, reaching DR 21 in a few months; using Bazzly to find niche Reddit posts to comment on and increase AI citations; and noting that some of his highest-ranking pages are product and case-study pages rather than blog posts, often cited by AI and ranking on Google. A quoted evaluation of Outrank said a SaaS site went from 0 to DR 64 in months using its listicles and guest posts, calling it the fastest way to jump-start a dormant blog and automatically get the first batch of DR 30+ backlinks, but said it requires manually editing top articles, adding real product screenshots, and cleaning up titles — and is “risky.” Tibo also said Outrank improves daily and is shipping many updates for users. Separately, he said that to be mentioned by ChatGPT you need to appear on Reddit, but Reddit is extremely hard to crack, warned against using a personal Reddit account because it will get banned quickly, and said the relevant tools are not cheap — via 1 2 3 4
2. Jack Ellis said he is copying 75 billion rows of data for database benchmarking. He tried to compare SingleStore Cloud with ClickHouse Cloud benchmarks, but ClickHouse auto-scales with queries and won; he said the result may also be due to schema design differences and does not plan to continue testing because it is too time-consuming. He also argued that selling only to indie developers is why software businesses have high churn — true even before AI, and now accelerating because indie developers love to self-build and self-host to save small amounts of money — and advised targeting enterprise customers instead — via 1 2 3
3. @levelsio said a common self-comforting mindset among software engineers is that “AI has not, will not, and will never affect me.” The quoted view argued that developers unwilling to admit reality often consider themselves “special” in technical interactions, and that their learned skills and experience have largely lost value unless they are among the 0.001% of developers working at frontier AI labs. He also said he now has plenty of time and believes a new leisure era is beginning: people direct work through conversation with agents, and agents run 24/7 in the background. He cited @felixrieseberg building a new homepage on iOS with Claude while cycling, and said he has run websites almost fully automatically with cron jobs for years, and that over the past year even adding features and fixing bugs has been fully automated by AI. He said the freed time can go to family, friends, entertainment, fitness, and the outdoors, with some into new projects, but his satisfaction from creative expression increasingly comes from outside the screen; he cited Keynes' prediction 96 years ago and argued that previous automation only pushed work to higher levels — from manufacturing to finance, management, and administration — and people chose to consume more and kept working, but this time “feels different” — via 1 2
4. @levelsio said data brokers are buying SaaS datasets for $100k to $2m, which can fill gaps when SaaS revenue declines; the data is then resold to AI training companies and frontier labs to integrate the SaaS's functionality into AI lab applications. The quoted post added that one reason more companies build software instead of buying SaaS is not wanting to be an unpaid data source, and said an M&A firm suggested Taptanium's legacy app codebase — a kind of training data — is worth about $45k per app because it predates the AI revolution and contains highly complex, correct hand-written concurrency logic. He argued this is inevitable and that in a sense Marx was right that capitalism will devour itself over a long enough time horizon; one way to protect yourself is to become a co-owner of frontier AI labs and companies, noting he holds large positions in Google, SpaceX, Nvidia, and others, and is considering buying OpenAI and Anthropic when they go public, though they may be overpriced by then, and said the benefit of the US stock market is that ordinary people can co-own these companies — via 1 2
5. @levelsio said Microsoft and Xbox are embracing the various web ports people make rather than issuing DMCA takedowns, calling it very difficult to do inside a big company facing an eager legal team, and thanked Satya Nadella and Asha Sharma for embracing creativity. The quoted post was an internal notice about the formation of XBOX LABS, mentioning that games have recently been quickly decompiled and remade for hardware they were not originally supported on, and that Windows PC versions have been ported to Apple Silicon or run in the browser via WebAssembly and WebGPU, with the latter now supported in nearly 90% of global browsers; it also said people spend 72 billion hours a year playing Xbox games, about five times the global time spent writing code. He also quoted a view saying many similar posts will follow but not to adopt that mindset because it is unproductive; it is still very early in the innovation adoption curve, and a better approach is to ask, “How do I make myself resilient to the change a year from now when someone can build my business in seconds with a prompt?” and act accordingly — there is still time to pivot, just do not stagnate — via 1 2
6. @levelsio explained how Hotelist shows real hotel ratings rather than inflated booking-site scores: for one hotel, booking sites showed a range of 4.2 to 5 and the hotel rating was 4.5; normalized within that range, it becomes 1.88/5. The quoted post said Airbnb ratings only range from 4.5 to 5 with a median of 5, most sites range about 4 to 5 with a median of 4.25, Yelp is the most honest with a real normal distribution, and Hotelist's own median is 3.5, with the goal of making most accommodations average/fair at 3.5/5. He also quoted a view that hotel “eco” messaging is likely a cover for cost control: if a hotel has 200 rooms and an average of 10% of guests hang a “do not disturb” sign, that could improve operating profit by 10% — via 1 2
7. Arvid Kahl said libgen and AA remain effective top-of-funnel distribution channels for his books: many people download and do not pay, but they would not have bought anyway, and some read and then buy, with thank-you emails as evidence — he called this “Freemium.” Tw93 shared a thread arguing that maintaining a “human feel” and uniqueness matters in the AI era: Chinese content on Twitter often attracts AI replies that feel like someone suddenly summarizing incoherently in the middle of a chat, disrupting the atmosphere, and such practices do not bring account revenue or data gains, and may instead get the account banned and waste tokens. The thread said the author likes to watch Testv videos while coding with AI, praising their human feel, refusal to take ads that deceive viewers, interesting product reviews, and records of life in Chongqing; by contrast, some top digital review scripts feel predictable in gestures, smiles, and filler words, which is tiring. It compared this fatigue to the flood of AI-generated images, videos, and articles: many would have seemed high quality before the AI era, but repeated cores become tiresome, and the author argues content can be non-repetitive and self-written, and even if quality is average it is far better than repetitive “high quality.” The thread acknowledged AI can greatly raise the level of work, letting people with ideas make things they could not before without asking for help, and said the author has benefited from model progress and large efficiency gains; but it argued products should incorporate one's own human feel, tone, and worldview rather than letting AI think for you. It worried that fully AI-ifying everything from idea to product to operations will lead to uniformity, and mentioned someone using abundant tokens to rewrite an engineer's years of product work, change languages, distill it in half a day, and upload it open-source to GitHub while belittling the original developer; the author thinks this may discourage engineers with ideas from continuing to contribute good code and ideas, and discourage creators with ideas from posting, ultimately letting quantity beat quality. The thread said that after experiencing both the pre-AI and current AI eras, the author insists on hand-typing all “just saying” content without AI generation, even with typos, awkwardness, or rambling, to preserve a human feel, and worries that full reliance on AI will make him dumber — via 1 2
8. Tibo shared a post saying someone applied three points from his newsletter about X — if a post performs poorly within 30 minutes, change the hook and repost; people bookmark content they will use later; long content beats short content — and changed a one-line post into a longer how-to guide, raising impressions from 1K to 30K and bookmarks from 2 to nearly 1K. He also shared that his newsletter is completely free and high value — via 1 2

5. Work, Society, and Personal Notes

1. @midwaydude said his son will take GCSE next year (noted as roughly equivalent to the mainland middle school exam) and he attended a prep lecture. The speaker said students often review five hours a day, six days a week in the final stretch, called that arrangement “terrifying” and likely to cause burnout, and recommended starting now with 1 to 1.5 hours a day and resting on weekends to ensure peak condition for the exam next May; the author was skeptical of both the “five hours is a bit little” framing and the speaker's advice — via 1
2. Nathan Barry posed a self-check question: if you had to give one last talk to the world, what would you say? If you cannot answer, you may not yet be doing what you are meant to do. He mentioned that Philip McKernan, author of One Last Talk, did a live coaching session with him on a podcast, bringing his anxiety about promoting his book from 5 to 0 — via 1
3. @levelsio said most of these services will be fully automated within a few years, with only the absolute high end keeping humans, ending the deterioration in service quality since COVID caused by labor shortages — which he linked to global low fertility and government payments making not working easier than working. The quoted post used a Bangkok motorcycle rental as an example: submitting documents through third-party KYC, choosing a bike online, paying with Stripe, and picking up and returning at the shop without face-to-face contact, communicating details only via WhatsApp; the downside is pickup must be after 14:00 and return by noon, unlike traditional rentals counted as full 24-hour days. He also said Barclays claims Tesla and SpaceX may merge before December 6 this year; if so, X would become the world's fifth-largest company, surpassing Amazon and trailing only Microsoft. He listed market caps as Nvidia $5.52tn, Apple $4.98tn, Alphabet $4.27tn, Microsoft $4.04tn, X $3.54tn (SpaceX $2.06tn + Tesla $1.48tn), Amazon $2.79tn, TSMC $2.23tn, Meta $1.87tn, Broadcom $1.65tn, Saudi Aramco $1.63tn, and said SpaceX is currently 7th and Tesla 11th. He also shared that Ben Wilson has stage 4 cancer and four children under 5; he invited Wilson on his podcast to talk about researching great leaders in history and what it feels like to be told at 38 that you will die, and called on listeners to help with medication, introductions, or donations. He also said he was shocked to see a “🌍 Be eco with us” card at a previously excellent hotel in Poland, having thought Poland had not been captured by eco/de-growth ideology, but it is fairly common even there, and he thinks it is hard to resist because Germany is next door and may be pushing it through paid promotion; he said Poland banned communism last year and argued the next step should be banning eco/de-growth, calling them essentially the same with the same ultimate goal. He said the main reply about Warsaw was “leave Warsaw and go to Krakow,” noted he is already there, and asked what to see and do in Krakow. He also posted, “Always bring her flowers, bring yourself Skyr.” — via 1 2 3 4 5 6

6. Tools and Product Notes

1. Ben Tossell gave feedback to @fayazara on Screendrop recording needs: when capturing by drag, he wants to be able to choose and set a default aspect ratio (for example, always show 16:9 with corresponding dimensions) and to support resizing from the corner to fit different screens — via 1
2. Tony Dinh said Steam has approved his game, with a planned release on November 1, and called for people to add it to their wishlists — via 1
3. Tibo said that if you want to be mentioned by ChatGPT you need to appear on Reddit, but Reddit is extremely hard to crack, warned against using a personal Reddit account because it will get banned quickly, and said the relevant tools are not cheap — via 1
4. @levelsio explained how Hotelist shows real hotel ratings rather than inflated booking-site scores: for one hotel, booking sites showed a range of 4.2 to 5 and the hotel rating was 4.5; normalized within that range, it becomes 1.88/5. The quoted post said Airbnb ratings only range from 4.5 to 5 with a median of 5, most sites range about 4 to 5 with a median of 4.25, Yelp is the most honest with a real normal distribution, and Hotelist's own median is 3.5, with the goal of making most accommodations average/fair at 3.5/5. He also quoted a view that hotel “eco” messaging is likely a cover for cost control: if a hotel has 200 rooms and an average of 10% of guests hang a “do not disturb” sign, that could improve operating profit by 10% — via 1 2
5. Daniel Vassallo said North American big cities look beautiful from afar but are disappointing up close — dirty, unsafe, noisy — and that many places in the region that are also beautiful “from the inside” are almost all outside cities — via 1
6. Arvid Kahl said he could watch a quoted piece of content for hours; the quoted post said its dystopian level is beyond what even Black Mirror predicted — via 1
7. Ben Tossell joked that “engineers were replaced, turns out designers are safe,” using sarcasm to express a judgment about AI replacing jobs — via 1
8. Ben Tossell said his Stanford talk is now online and mentioned he woke up at 2 a.m. — via 1
9. Jon Yongfook said only about 8 real people are on X and the rest are “clankers” — via 1