Hey, it’s Nico.

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

Here are the 3 most important things of today’s issue:

• Ollie, an AI family assistant, shut down — find out why below

• Why AI could make it viable to build software for one person or team

• In an MIT experiment, one AI rewrote another’s weights without retraining it. — learn why this matters below

Tell Laura what your business needs. Watch it get made. ^AD

For 25 years, I bought professional services, most recently as Chairman and CEO of Hitachi America. The world’s best expertise was available only to the few companies that could afford it, so I built Kaldune to make it accessible to every business.

You tell Laura, our AI creative director, what you need in plain English. She asks only the questions that you can answer, brings in the right specialists, and lets you watch them do the entire job in real time: a market study your board can act on, a brand and its launch film, or a working app for your team.

See what the studio has created, then simply tell Laura what you need.

Try Laura →

This Week In Startups

🔗 Resources

How 1,014 software buyers responded to three pitches for the same AI product

How PostHog handed off six engineering chores to scheduled agents, with prompts you can adapt

Why AI could make it viable to build software for one person or team

Three paid ad playbooks for B2B teams

📰 News

Meta acquires Conductor, a tool for running multiple coding agents in parallel

OpenAI releases Images 2.6 with support for creating GIFs

Claude can now analyze your spending with its new Personal Finance tool

OpenAI is finally merging ChatGPT’s Chat and Work tabs

💸 Fundraising

Zipline, a startup that operates an autonomous drone delivery service, raised $1.5B in funding

Fal, a startup that runs AI models that generate images and video, raised $1.3B in Series F funding

Sentra, a startup that helps companies find and protect sensitive data across cloud services, raised $150M in Series C funding

Pliant, a startup that helps banks and software companies issue business cards and manage spending, raised $80M in Series C funding

Fail(St)ory

The Family Group Chat

Ollie was an AI assistant you could add to the family group text. It kept track of schedules and reminders and could help plan dinner in the same thread. The company says it will close the assistant and its separate meal-planning app on November 6.

Ollie asked families to pay for help keeping everyone in the loop. Doing that well meant staying connected to the apps they already used, some of which were becoming the home for competing assistants.

What Was Ollie:

A change to the school schedule can leave one parent checking the calendar, figuring out who can pick up the kids, and texting everyone who needs to know. Ollie’s pitch was to take over the work of spotting those changes and keeping the family up to date.

One parent could add Ollie to a group text with a partner or caregiver and give it access to an inbox and calendar. If the school emailed about an early dismissal, Ollie could pull the new pickup time from the message and include it in a briefing for the whole group.

People in the thread could also ask Ollie to set a reminder, add something to the grocery list, or help plan dinner.

Ollie had already built a meal planner that turned a week’s dinners into a shopping list. The company says families asked for help with other jobs around the home, which led it to build the text assistant.

Households could start for free, and one paid subscription covered the family thread. The paid plans came with daily and monthly allowances and the option to buy extra replies.

Behind a single text, Ollie might need to pull information from an inbox, use an AI model to work out what mattered, and send the result through its messaging provider. Subscription revenue had to cover those services and the work of keeping the connections running.

For families to keep paying, Ollie had to get those small jobs right week after week. It also had to give them a reason to pay for a separate assistant as Google began testing shared briefings and task lists for families.

The Numbers:

• 📅 Founded in 2023

• 💰 Reported seed round of $7.5M

• 👥 Team of eight in August

• 🍽️ Meal app reach of 90,000+ families, according to Ollie

• 💳 Paid assistant plans at $25 or $100 a month

Reasons for Failure: 

• Other assistants moved into the same chores. Meta made most everyday uses of Muse free. Google began testing shared briefings and family task lists inside tools households already used. Each gave families another way to get some of the work Ollie charged for done. Ollie's CEO named Muse and Instinct when he said competition had made it harder to stand out, grow, and raise money.

• Ollie had to become a household habit. One parent could add Ollie to the group chat. For a job like school pickup, the thread had to become part of how the family actually made plans. If a partner kept coordinating elsewhere, the parent who set up Ollie could still end up relaying updates and confirming who was picking up the kids. That may have made it harder to turn one person's trial into a routine the family wanted to keep.

• The $25 upgrade was a hard sell. Ollie’s paid plan raised the daily refill from six replies to ten and the monthly pool from 50 to 150. Instinct was free for invited beta users. For families who could get similar help from Instinct, $25 a month was a lot to pay for a higher reply limit.

Why It Matters: 

• Watch what people do when the free allowance runs out. If they can wait until tomorrow or finish the job with a free competitor, test whether they’ll pay to have a recurring job done on time.

• A family assistant can have an active user and still leave one parent doing all the coordinating. Track how often a second person acts on the assistant’s updates.

Trend

When AI Rewrites AI

In an MIT simulation, a car kept driving straight at a ramp and getting stuck. The neural network driving it failed all 256 test runs. MIT gave GPT-6 Astra a description of the scene, a recording of one attempt, and the 171,664 numbers that made up the controller’s weights.

Astra rewrote some of those numbers in one reply. The car backed up to give itself room to accelerate, then cleared the ramp in all 1,024 tests of the edited controller.

Weights are the numbers a model learns during training. They are the model’s brain.

This week’s experiments gave an AI a behavior to change and access to another model’s weights. In several cases, the AI figured out which numbers to rewrite and changed them directly, without retraining the model.

Why it Matters

• Changing a few weights can help reveal which parts of a model drive a specific behavior.

• Some model failures could be repaired without another training run.

• An AI might eventually work out an edit to its own weights and run a new copy with the change.

What the Edits Changed

In another experiment from the same report, MIT gave Astra a small Qwen language model and asked it to improve its math by editing the weights. Astra used Python to inspect the model’s internals and ran tests on practice problems.

It traced Qwen’s tendency to repeat itself to a signal in two layers and weakened that signal by changing 0.016% of the weights.

On a math benchmark Astra hadn’t seen during the experiment, Qwen’s score went from 7.9% to 48.1% under the same settings.

In a separate experiment, Valerio Dolci used Claude to edit the weights of a 12-billion-parameter Gemma model. He wanted Gemma to use a particular Italian tense when translating an action that was still underway at a point in the past. So, he wanted a very specific change in the model.

Claude found a signal that distinguished those cases from actions that had already ended, then rewrote a few neurons to make Gemma use that signal when choosing the tense.

On a fresh set of sentences about actions still underway, Gemma used the less preferred tense 16% of the time, down from 32%. The edit also changed 18 of 200 unrelated translations, a reason to test the rest of the workflow before switching models.

The Trend

MIT calls this weight-space programming.

Human researchers had already used direct weight edits to suppress repetition. The problem is that it was extremely hard and time consuming to do.

What’s new in these experiments is who figures out the edit. Given a behavior to change and tools to inspect the model, a general-purpose AI can sometimes trace that behavior to a small part of the network and change the weights there.

Could an AI do this to its own weights? In a separate experiment, a coding agent fine-tuned the saved model that future instances of itself would load. I’d like to see the direct version of that experiment. Give an agent a copy of its own weights, let it work out an edit, then start a new instance from the edited copy. See whether it can pass a fresh test for the behavior it set out to change and still do the work of finding the next edit.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico