Supabase Acquires Turso Database as Agentic Development Practices Mature

Key Takeaways

1. Supabase has acquired Turso Database, with more details promised at a keynote within hours.
2. A developer shares a detailed agentic development workflow: adversarial review at every step, TDD embedded in the agent loop, and context maximized into the codebase.
3. Karpathy outlines how LLMs can generate custom explainer videos, interactive HTML pages, and diagrams, shifting human work toward supervision and understanding.
4. A Mole developer manually emailed every unactivated buyer since launch through September, revealing gaps between purchase and actual use.
5. A launch platform operator observes 300+ daily submissions, mostly low-quality and abandoned quickly, but sees opportunity in underserved non-technical niches.
6. EU bankruptcies hit their highest level since records began in 2019, with new business registrations stagnating or declining.

1. Supabase Acquires Turso Database

1. Supabase has acquired Turso Database, according to a repost by Ben Tossell, who said he is a long-time supporter of the Turso team and product and that the two sides have major plans to disclose at a keynote within hours. Tossell also mentioned attending the Supabase Select event and noted that a PlanetScale truck outside the Supabase event did not stay long. The acquisition details and strategic rationale remain undisclosed pending the keynote. — via 1 2 3

2. Agentic Development Workflow and LLM-Assisted Understanding

1. Arvid Kahl summarized his agentic development process: every step should include adversarial review, ideally with a third agent acting as arbiter; test suites must cover unit through end-to-end with no exceptions, and TDD must be embedded into the agent workflow; non-code parts, especially architecture and decision stages, rely more on human review. He also argued for maximizing context into the codebase — documentation, ARDs, customer ICP descriptions, runbooks, and schema annotations should live inside the repo rather than as attachments — and for tracking every agent conversation and change (e.g., via Entire or similar tools) attached to tickets or PRs. — via 1
2. Kahl further argued that development and test database contents matter more than ever: dry runs, workflow simulations, edge-case investigation, and performance measurement should all happen before production, and seed data is not fake data — even if smaller in volume, its shape must represent production. He also proposed that code best suited for agents may not look best to humans, and that humans should position themselves as verifiers rather than proofreaders, noting that comprehension is now a function of tools rather than a mental state — a shift he called novel and somewhat frightening. — via 1
3. Kahl reposted content questioning the evidence for adversarial review's effectiveness: whether different models adversarially reviewing each other makes a difference, how the same model responds to requests for improvement within the same session versus across sessions, and who is doing this research — asking whether he would have to do it himself. He also reposted a view that wandering in World of Warcraft while watching agents work is similar to a real walk, in that both distract the conscious mind and let the subconscious process other things. — via 1 2
4. Karpathy proposed multiple ways to understand LLM output: using the ASD-STE100 controlled language specification to have LLMs explain content for improved readability, or requesting "80% close to ASD-STE100" to relax strict constraints. He suggested having LLMs generate diagrams and interactive HTML web pages, and said he is most bullish on generating fully customized explainer videos on any topic (e.g., 3b1b style, with ElevenLabs API voiceover), calling this direction already feasible. He summarized that as LLM capabilities improve, they will autonomously take on more foundational work while human work rises to supervision and understanding; with intelligence and code increasingly abundant, one can request large, customized, disposable software artifacts — such as web apps and video explainers — that were previously infeasible. Tibo reposted this, saying he previously did not understand how Shazam works and believes customized explainer videos will transform education. — via 1

3. Product Operations and Market Observations

1. Tw93 manually emailed every unactivated Mole buyer from launch through the end of September, resending license keys by month and country language, emphasizing this was not a promotion or activation nudge but a reminder. He received replies from users in multiple countries: some bought to support development or prefer terminal/CLI versions; some bought two copies for personal use and as a Christmas gift for family; some are waiting for a new Mac because their company bans unlicensed software; and some bought during a discount but do not yet have a MacBook. Some users entered incorrect email addresses (e.g., gmai, con, cloud) without contacting support, and some repurchased because they did not receive the first email. Users who did not receive activation codes are a very small share of total users, and many were resolved through daily support emails. Tw93 argued there is a gap between purchase and actual use — users do not necessarily "see → buy → use immediately" — and developers should proactively take an extra step rather than assuming the relationship ends after payment. He noted that engineer-turned-product people tend to obsess over product features while neglecting purchase-flow data and experience optimization around non-product functions, and advocated doing "unscalable" things in an era of lowering code barriers: manually walking through the flow to find optimization points. He plans a new Mole feature: after payment, users are automatically redirected to a website activation page, the payment platform sends a synchronized email, and clicking activate on the web page can one-click launch the Mole client and auto-fill the key, addressing incorrect or blocked emails. — via 1
2. Jon Yongfook reposted and commented on an article arguing that "fast shipping is dead, the new rule is slow thinking," agreeing with all its points but saying the game continues with changed rules. The quoted post, from a launch platform operator receiving 300+ product submissions daily, offered two observations: everyone is building the same things and creativity is dead for many, which paradoxically means opportunity to stand out; and barriers to entry are nearly nonexistent, but average product quality is extremely low, with most built in a few hours and abandoned the next day. The poster argued that noise makes founders despair, but removing noise changes little — many underserved niches remain, mainly for non-technical audiences. The effective approach is to spend time, pick one thing, and build something genuinely useful, citing an example of someone with years of experience in the amusement park and museum industry who built SaaS for store management chaos at such venues, profiting through high-ticket subscriptions, in-person demos, and staff training rather than UGC content or cold email blasts. — via 1
3. Levelsio noted that EU bankruptcies have reached their highest level since records began in 2019, with new business registrations stagnating or declining. He noticed this after two interior design stores in the Netherlands went bankrupt shortly after receiving goods, and argued that European regulation and taxation suppress existing operations and entrepreneurial willingness. — via 1

4. Tools, Products, and Personal Observations

1. Ben Tossell reposted content saying Factory can automate low-leverage work, with its automatic model routing removing the burden of picking the best among 10 new models released that week. He also said personal agents should proactively build apps and websites for users. He criticized a product experience, saying it may be great for some but not good enough for people like him who are "non-technical but not ordinary users," and that the problem should not be blamed on users or skill gaps but fixed in the product, calling the experience unhelpful to its narrative. — via 1 2 3
2. Jack Ellis said he will build a to-do app using 62,500 free ChatGPT credits. He had Opus 5.5 attempt to break his new AI bot and said one of the messages he received showed Anthropic has no restraint. When asking a friend at a successful company who heavily uses AI for development about building versus using SaaS, the friend's reasoning was: you need to pick products and bear selection responsibility, so you should choose well-regarded SaaS. — via 1 2 3
3. Jon Yongfook is looking for a Windows machine for development that is not too painful to use, noting his team is now building desktop apps. He added a Mac virtual camera app that blurs faces, supporting full blur or keeping only yourself clear, compatible with Zoom, Google Meet, and other video apps, and free. — via 1 2
4. Daniel Vassallo compared F1 rules to question German autobahn unlimited speeds: in F1, cars more than 7% off the fastest qualifying lap are not allowed to race, yet on the highway 75 mph vehicles share the road with 175 mph vehicles, which he considers a major threat to others' safety. — via 1
5. Levelsio said he chose a Tesla Model Y for safety, considering it one of the safest cars with the highest survival rate, and does not care whether it looks pretty or cool. He acknowledged the Model Y looks like an ordinary sedan lacking personality but said he does not mind; he thinks the Cybertruck has character, and a friend's Jimny looks good but is neither fast nor safe. He cited a ranking based on 2026 IIHS crash tests and independent composite scores where the Model Y (2026/27) ranked first, followed by Genesis GV80, Lincoln Nautilus, Infiniti QX60, and BMW X3, most earning Top Safety Pick+. — via 1 2 3
6. Levelsio reposted content saying the Prometheus statue has been cast in Paris, shipped across the Atlantic to Starbase, and assembled, taking 100 days from zero to completion, and that construction of a "most magnificent" new city will begin within months. He also made coworking meetups a separate category with a new "regular meetup" option for organizing in any city worldwide; the quoted content said 70 weekly coworkings have been organized along with movie nights, art nights, and barbecues, arguing that proactive organizing is key to preventing social shrinkage after 30 and that active socializing matters as much to health as physical fitness. — via 1 2
7. Levelsio found V60 coffee made with reverse osmosis water tasted flat because TDS was only 25 after adding a remineralizer, versus about 60 with plastic bottled water. Using Lotus Coffee's calcium, magnesium, potassium, and sodium drops, he raised TDS to 156 and the flavor improved noticeably, but he is considering making his own solution or switching to a remineralizer that automatically reaches 150 TDS. — via 1
8. Nathan Barry said his podcast has performed well recently: an episode with @jasminestar got 184,000 views, a historical high, and the next episode with @theChrisDo also started well. He said he is having more fun than ever. — via 1
9. Ben Tossell recalled reuniting with early customer Sharath, who said he paid $99 in 2018 for a NewCo bootcamp, learned no-code building and product shipping, has since shipped many products, sold some, and entered the tech industry; Tossell said he is still not good at naming companies. He reposted someone who ran into him at a San Francisco event about seven years after nocode_conf, saying that without that connection they would not be teaching today, and that friends made online and offline are what matter. He also reposted a view that the person does not know anyone with a net worth over $1 million who drinks (or at least gets drunk), calling it an interesting correlation; he and his wife stopped drinking about five years ago when they became pregnant. — via 1 2 3
10. Tibo said he is extremely proud of his products, almost not caring whether users buy them because he simply likes them. — via 1
11. Jon Yongfook said he posted on LinkedIn and commented "this is the hell that AI brings." An unnamed project's first-day data: 573 visitors, 1.2k page views, MRR of zero because the product is free with no paid option. — via 1 2
12. Ben Tossell said hackable badge merchandise is great. — via 1