Key Takeaways
- Arvid Kahl argues that over-compressing prompts backfires because models need context to stay clear.
- Tw93 warns that browser-automating agents on important accounts can get users banned, and recommends official APIs or CLIs.
- Levelsio reports finally porting Urban Terror to the web via WASM after two failed attempts, and reflects on AGI-era real estate.
- Jack Ellis finds ChatGPT appends UTM parameters while Claude does not, and is adding an AI traffic tab.
- Tibo shares an AI-generated explainer video workflow and a 30-day AI explainer series.
- Daniel Vassallo notes Anthropic still uses Typeform, questioning the "SaaS is dead" narrative.
- A Hong Kong luxury project faces construction-quality allegations, with only a property self-media outlet reporting it.
1. AI Agents and Automation Risks
- Tw93 reports that a friend's GitHub account was banned, possibly because an automated Agent simulated browser operations in violation of official rules. He recommends hard rules for Personal Agents: do not use browser automation on important accounts, especially X, Reddit, Xiaohongshu, GitHub, and Gmail; prefer official APIs or official CLIs; inspect the underlying implementation of MCP and third-party CLIs, since some merely wrap web simulation, cookies, or private interfaces; and if no suitable official interface exists, have the Agent stop and ask the user to handle it manually rather than attempting workarounds that could trigger risk controls or bans. He distinguishes this from the more mystical "avoid Claude ban" advice, noting many such articles contain affiliate links and are often selling cloud hosts or services. — via 1
- Arvid Kahl argues that prompt compression actually reduces effectiveness, because models need more context to achieve clarity; after compression, the result is at best a restatement, a rephrasing of the idea in multiple ways, or a drift off-topic beyond the original scope. — via 1
2. AI Products, Traffic, and Tooling
- Jack Ellis observed that ChatGPT appends UTM parameters while Claude does not, and showed the human traffic these referral channels bring to websites. He decided to add a dedicated AI tab to his dashboard and integrate the insight into his AI chat feature. He also forwarded a note that his team had long wanted a tab to see AI visitor sources, since they previously only knew about ChatGPT and Claude and were unaware of other sources. Separately, he filed a Cursor feature request for "pinning skills," because he commonly uses about five skills but sees over 100 when typing "/", and forgets the ones he creates. — via 1 2 3
- Tibo shared that he had Squad use Revid CLI to make a two-minute video on how human work has changed over the past 300 years; according to his description, the tool did the research, sourced visuals from the internet, produced the soundtrack, and handed everything to Revid to generate the final result. He also launched a 30-day series in which AI explains one everyday thing that is hard to explain, starting with JPEG, and he admitted he thought he understood JPEG but did not. He cited an "explain how Shazam works" example, said he previously had no idea, and agreed with Karpathy's related view that customized explainer videos will change education. — via 1 2
- Daniel Vassallo noted that Anthropic still uses Typeform for email marketing and customer research, and the forwarded post argued this shows some SaaS solutions are not dying as many predicted. He questioned what it means that a frontier lab with the strongest AI models chooses Typeform rather than quickly building a rough in-house solution. — via 1
3. Web, Games, and Real Estate
- Levelsio reported that after failing six months ago to port Urban Terror, which is based on the Quake 3 engine, to the web, he tried again with Opus 5.5 and still failed. He finally used Fable 5.1 in high-effort mode to locate the issue: the engine struct was 4 bytes larger than the UT bytecode expected, causing entity counts and list reads to be offset by 4 bytes. After rewriting the bytecode, it succeeded, and the game now runs 100% in the browser via WASM with multiplayer servers. — via 1
- Levelsio argued that AGI may polarize society: one group absorbed in AI content and government-provided basic income, another turning offline toward family, local community, and health, forming an elite counter-movement often living in small communities. He admitted he leans toward that lifestyle, and mentioned @okaythenfuture thinks it will happen in city centers, possibly both. He stressed that in a post-AGI era most goods are nearly free while real estate and land remain scarce, so he now considers buying property and land reasonable. The quoted post argued AI slop and realistic AI companions will flood the internet, spurring an elite counter-movement against being overly online and pushing up demand for core locations in global first-tier cities, naming central Tokyo, Bangkok, Manhattan plus Brooklyn, London Zones 1/2, and prime Paris areas, and calling the 2030s the most interesting urban decade. — via 1
- A Hong Kong luxury seaside project, One Stanley in Stanley, was revealed to have construction-quality problems: rebar quantity was 50% less than the drawings, and after owners chiseled open the floor they found rebar arranged randomly, seemingly without tying work, and with many breaks. The Buildings Department's demolition inspection also determined it was not built according to the drawings, but said there was no obvious structural danger. Owners and the developer are still in dispute. The author questioned what this means for HK$10 million to HK$100 million homes when a HK$100 million luxury property is like this, and cited someone saying that in Hong Kong you should only buy buildings over 20 years old, otherwise you can only "laugh secretly." The author noted the only media outlet reporting this was a property self-media account. — via 1
4. Brief Signals
- Jon Yongfook noted that by 2026 AI capabilities have improved greatly, yet everyday bank transfers still require screenshotting a QR code and uploading it to the banking app, a process that remains cumbersome. — via 1
- Nathan Barry said a personal brand needs to be clear enough that the audience can sum up what you stand for in a few words, otherwise the brand is not strong enough; Chris Do believes every personal brand needs three things: a cause, an enemy, and an action slogan tying the two together. — via 1
- Ben Tossell said his first tech job was working with someone he calls the "GOAT." The account also forwarded a view that all testing frameworks should remove queues, just as plan mode was removed. — via 1 2
- Levelsio said he now builds things faster than he generates new ideas, which is a strange experience, and forwarded content about "the death of software." He also commented on Qatar's startup support policies, calling it another example of a city-state attracting European tech talent with better living, working, and regulatory conditions. — via 1 2 3
- Jack Ellis quoted another tweet and said Astra 6.1 is confirmed for release next week. He also criticized the narrative that "people with higher net worth don't drink," pointing out some rich people drink heavily, some drink less, and some not at all, and one should not look for absolute conclusions. — via 1 2
- Tibo shared Nikita Bier's method for building viral apps. — via 1
