Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Presearch, a startup building a decentralized search engine, has shut down — learn more below.

• How much can you delegate to agents?

• Kimi K3 is out and everyone is debating about open-weight models — learn more below

This Week In Startups

🔗 Resources

Steal this data from Anthropic's 1 million user interviews

Why every marketer needs a GitHub

The next AI moat isn’t a better model

How much can you delegate to agents?

📰 News

Friend, the lonely AI wearable, returns with a new voice and a much bigger price tag

Google’s AI search is rapidly becoming the default

LinkedIn adds a button to report AI-generated ‘slop’

DoorDash is building its own drone delivery business

💸 Fundraising

Sila, a silicon-carbon anode materials startup, raises a $300M private equity round 

Neo, an enterprise platform for securing and governing AI agents, raises $100M

Twenty, an offensive cyber warfare startup, raises a $30M follow-on investment

Natural, a payments infrastructure platform for AI agents, raises a $30M Series A 

Fail(St)ory

Who Owns Search?

Presearch tried to build a private, community-run search engine and pay people to use and operate it with crypto tokens.

It shut down on this week, after nine years, caught between the cost of running search infrastructure and an audience too small to fund it.

What Was Presearch:

Presearch wanted to build a search engine that no single company controlled. Users would get more privacy, community members would help run the infrastructure, and the PRE token would reward both groups.

Its first product was a search dashboard. You entered one query and chose whether to send it to Google, DuckDuckGo, Wikipedia, LinkedIn, or dozens of other services. Most users kept using the default, so Presearch moved toward a normal search experience with one query and one results page.

Presearch did not crawl most of the web itself. It gathered results from other search engines and data providers, combined them, and displayed them under its own interface.

To process those searches, Presearch used computers run by members of its community. These computers, called nodes, received anonymized queries, fetched results from outside services, and sent them back to Presearch.

The goal was to spread the work across many independent operators instead of relying only on company-owned servers. This could lower infrastructure costs, make searches harder to link to individual users, and reduce the amount of control held by one company.

Anyone could run a node by installing Presearch’s software on an online computer with a public IP address. Operators staked PRE tokens and earned rewards based on the number of searches they handled, plus their speed and reliability.

Users also earned small amounts of PRE for searching, giving them a reason to switch from Google. Advertisers were supposed to fund the system through keyword ads, display ads, paid ad-free search, an AI subscription, and an API.

Presearch later started building its own index, called Indee, focused on forums, small publishers, and creator sites. The full plan was a private search engine with its own index, infrastructure supplied by its community, and tokens paying the people who used and operated it.

The Numbers:

• 🚀 Founded: 2017

• 👥 Monthly active users: Around 115,000

• 🔎 Daily searches before shutdown: Nearly 500,000

• 🖥️ Active community nodes: More than 38,000

• 💵 2024 revenue: $208,547

• 📉 2024 net loss: $267,414

• 🏦 Cash at the end of 2024: $23,735

Reasons for Failure: 

• Rewarded searches were hard to monetize: PRE tokens gave people a reason to search, but they also attracted bots and users chasing rewards. Presearch spent a lot of money checking traffic that advertisers might not value.

• Search needed much greater scale: Presearch was close to 500,000 daily searches and still could not cover its costs. Management believed it needed one million legitimate searches per day, with no clear path to reach that level fast enough.

• The system stayed expensive to run: Community nodes handled some work, while Presearch still managed routing, fraud checks, external data sources, ads, accounts, and the core search product. Its own index added another large technical cost.

• Funding ran out first: Presearch ended 2024 with $23,735 in cash and raised only $183,328 through its 2025 crowdfunding campaign. New products and enterprise deals were still under development when the company needed immediate revenue.

Why It Matters: 

• Pay users only for behavior that creates clear economic value, because cheap actions like searches are easy to fake and costly to verify.

• Decentralized infrastructure can increase costs when the company still handles routing, quality control, fraud, support, and monetization.

• Model the exact usage level needed for break-even before scaling incentives, then prove revenue per user at a smaller size first.

Trend

Open-Weight Week

This week, everyone has been talking about open-weight models. 

The long-awaited Kimi K3 finally released its downloadable weights, Nvidia backed a major industry letter, Anthropic published its objections, and Mark Zuckerberg joined the debate with an op-ed of his own.

Why it Matters

• Open weights shift control from model providers to startups. You can run models yourself, keep data in-house, and adapt them—but you also take on infrastructure, security, and maintenance.

• Most companies won’t run K3 directly. At 2.8 trillion parameters, it’s too large for most teams, which means value moves to hosting, inference, and model compression layers.

• Every major player has a different incentive. Nvidia wants more compute demand, startups want flexibility, and frontier labs worry about losing control once weights are released.

Kimi K3

Moonshot AI launched Kimi K3 on July 16 and released the weights on July 27. It is the largest open-weight model released so far: 2.8 trillion total parameters, with 104 billion active at a time, native vision and a one-million-token context window.

K3 can handle long coding jobs, agent tasks, research, reasoning and work across text and images. Moonshot’s tests place it among frontier models, though still behind the strongest closed systems. The release drew so much attention because nobody had made a model this large available to download before.

Its mixture-of-experts design only activates a small part of the model for each token. That makes inference far lighter than running a dense 2.8-trillion-parameter model, though serving K3 still requires serious infrastructure.

The Controversy

While everyone was discussing Kimi, many of the biggest AI companies started debating the role open-weight models should have in the future of AI.

On July 24, Nvidia, Microsoft, Meta, Hugging Face, Mistral and more than 20 other companies signed a letter supporting open weights. OpenAI, Google, Amazon and others joined soon after. They argued that downloadable models give companies more choice, let them run AI on their own infrastructure and reduce dependence on a small group of model providers.

Nvidia then launched the Open Secure AI Alliance, which focuses on using open models for cybersecurity. The company pointed to a recent Hugging Face investigation where proprietary models refused some forensic tasks, while an open model could run locally and review more than 17,000 suspicious actions.

Anthropic published the most prominent counterargument. It supports open releases when the risks remain limited, but wants stronger testing as models become more capable in areas such as cybersecurity and biology. Once weights are released, safeguards can be removed, usage becomes hard to track and the model cannot be recalled.

The debate now comes down to where companies draw that line. Nvidia, Meta and Microsoft see broad access as useful for competition, security and customer control. Anthropic wants more evidence before the most capable models can be downloaded by anyone. Kimi arrived at the exact moment the industry was being forced to take a side.

The Trend

This week was a reminder that a large part of the AI race is happening among companies most people in the West barely follow. Moonshot turned Kimi K3 into one of the week’s biggest releases, then watched American tech companies spend days arguing over what models like it mean for competition, security and control.

These releases keep widening the field. Every capable open-weight model gives developers another serious option and puts more pressure on the major labs to explain what they will release, what they will keep closed and where they draw the line.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico