Hey — It’s Nico.

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

If you only have one, here are the 5 most important things:

• Astra, a startup making AI for account executives, shut down — learn why below.

• What's Next For Design Tools

• X plans to introduce ads in Grok’s responses

• Anaconda raised $150M for Python AI tools.

• OpenAI just released two open-source models — learn more below.

This Week In Startups

🔗 Resources

Context Engineering: Building Production-Grade AI

What's Next For Design Tools

Linear’s Path to Product-Market Fit

Failure Modes for Engineering Team Leads

📰 News

X plans to introduce ads in Grok’s responses

Upwork is buying its way into corporate staffing beyond freelancers.

Google denies AI search features are killing website traffic.

ElevenLabs launches an AI music generator, which it claims is cleared for commercial use.

💸 Fundraising

Aerospace startup Jeh Aerospace raises $11M

Anaconda raised $150M for Python AI tools.

AI-powered fintech Alaan raises $48M

Fintech startup Rillet lands $70M

Fail(St)ory

AI for AE

A few months ago, AI-focused SaaS startup Astra was making waves. It had just secured backing from Perplexity’s founder Aravind Srinivas and was positioning itself as a game-changer for sales teams.

Now, less than four months after that funding round, Astra has shut down.

What Was Astra:

Founded in 2023, Astra pitched itself as the “Chief of Staff for every account executive.” The vision was to replace the busywork that eats up a salesperson’s week with an AI-powered assistant that could handle it automatically.

The platform promised to automate up to 80% of an account executive’s activities — from mapping sales territories to tracking lead conversion ratios — while connecting seamlessly with enterprise tools like Salesforce, Slack, and Google Drive. 

The idea was that AEs could stop spending hours updating CRMs or chasing documents and instead focus entirely on closing deals.

It was a compelling pitch. The founder claimed the platform was so unique it had “no direct competitors” and was able to land two enterprise clients during its beta phase. The founders believed they had found their wedge into the crowded AI space, and with high-profile backing, the future looked bright.

But Astra never made it out of beta. Four months after announcing its funding, the lights went out.

The Numbers:

• 📅 Founded: 2023

• 💰 Funding: Backed by Perplexity AI founder Aravind Srinivas

• 🤝 Clients: 2 enterprise customers onboarded during beta

• 🚪 Shut down: July 2025

Reasons for Failure: 

• Co-founder Misalignment: In his LinkedIn post announcing the shutdown, CEO Supreet Hegde hinted at disagreements with co-founder about the pace of growth. It’s not unusual for early-stage startups to hit points of tension between founders — but when visions diverge early, it can paralyze decision-making at the exact moment the company needs clarity and speed.

• The Enterprise Sales Trap: From day one, Astra went after large enterprises. Bigger contracts were the lure, but so were long, slow sales cycles. As Hedge explained: “Working with larger companies meant navigating lengthy sales cycles, especially as an early-stage startup asking clients to trust us with sensitive data”. Convincing a Fortune 500 company to hand over access to sensitive internal systems is a drawn-out process — and Astra didn’t have the runway to wait it out.

• Market Confusion in AI Agents: The AI-agent market is crowded and chaotic. Many companies are curious, but don’t know how to evaluate or trust these tools. Hegde admitted the current hype cycle made things harder: “The surge of interest and confusion surrounding AI agents added yet another layer of complexity, with many clients unsure of whom to trust or how to evaluate these AI agents.”

Why It Matters: 

• Early traction isn’t momentum — two beta clients aren’t enough without a launch-ready product.

• Big customers can be a slow death — long sales cycles can quietly kill early-stage startups.

• Co-founder alignment is oxygen — lose it, and everything else becomes irrelevant.

Trend

OpenAI Finally Goes Open-Source

Last week, I talked about how Chinese AI companies like Kimi K2 and Z.ai were racing ahead in open-weight models —and how OpenAI seemed to be sitting on the sidelines.

Well, they just proved me wrong.

After years of keeping their best work behind closed doors, OpenAI finally stepped into the open-source arena with the release of two reasoning-focused models: gpt‑oss‑120b and gpt‑oss‑20b. And they didn’t just dip their toes in—they made these models available under the Apache 2.0 license, meaning you can download, run, and even modify them locally.

This is a big shift for OpenAI and potentially a huge moment for the AI ecosystem.

Why It Matters:

• Puts OpenAI back in the open-source game — After being criticized for years for locking up their best models, OpenAI can now compete directly with Meta’s Llama series and Chinese open models.

• Levels the playing field for startups — Founders can now run top-tier reasoning models on their own hardware, avoiding cloud costs and keeping sensitive data in-house.

• Signals a possible industry pivot — If even OpenAI is opening up, other major players might follow—or risk losing developer mindshare.

What They Released

OpenAI launched two models:

• gpt‑oss‑120b — Near parity with OpenAI’s own o4‑mini on reasoning benchmarks, but optimized to run on a single 80 GB GPU.
  

• gpt‑oss‑20b — Similar performance to o3‑mini, but lightweight enough to run on a 16 GB device—perfect for edge applications and local deployments.

Both are transformer-based models using a Mixture‑of‑Experts (MoE) design. Instead of activating all parameters for every request, they only “wake up” a subset of parameters. This drastically reduces compute cost without sacrificing performance.

Like all other OpenAI models, the open-source models support structured outputs, tool use (think Python execution), and web search. The models can also scale their reasoning depth depending on task complexity.

How This Changes Everything

OpenAI has resisted releasing open models for over six years, citing safety risks. But a lot has changed:

• The competition has heated up — Meta’s Llama models dominated the US open-source space. Meanwhile, Chinese players like DeepSeek, Kimi, and Qwen have exploded in popularity.

• Developers want control — Startups and enterprises alike have been gravitating toward open models for cost savings, privacy, and customization.

• Pressure from within — Even Sam Altman recently admitted OpenAI had been “on the wrong side of history” by holding back open models.

In many ways, this release is more than just a product launch—it’s a signal. OpenAI is showing that it wants to remain relevant not just through proprietary models, but also by engaging with the broader open-source community.

It’s a reminder that open innovation still has a place, even in an industry increasingly dominated by closed systems and billion-dollar cloud contracts. If this trend continues, we could be looking at a much more decentralized, developer-first AI ecosystem in the near future.

Help Me Improve Failory

That's all of this edition.

Cheers,

Nico