Spotify stopped coding

Welcome to AI Toast!

Here’s what’s on the menu today:

• Spotify stopped coding 🤯

• Google’s deep think gets a massive upgrade

• Superhero Tools for research

• Secure your AI Agents with Teleport’s Agentic Identity Framework

• Quick Toasts: Elon’s new Moonbase

Total read time: About 5 minutes, perfect for a quick music brea

Spotify Top Devs Stopped Coding In December

Instead of manually coding, Spotify’s developers are now prompting AI to build and fix features for them.

• Spotify co-CEO Gustav Söderström shared that engineers are now heavily relying on AI to build and ship features.

• The company shipped over 50 new features and updates in 2025, accelerating product development.

• Engineers use an internal system called “Honk”, powered by generative AI and Anthropic’s Claude Code, to write, fix, and deploy code.

• Developers can prompt Claude from Slack even during their commute to fix bugs or add features.

• Updated app versions are sent back via Slack for review and can be merged before the engineer even reaches the office.

• Spotify says this workflow has dramatically sped up coding and deployment.

• The company is also building a unique music-related dataset that LLMs can’t easily replicate.

• On AI-generated music, Spotify allows metadata disclosures while continuing to monitor spam.

This signals a shift from AI as a coding assistant to AI as a primary builder inside product teams.

Personal Take: I think the most interesting part isn’t that AI writes code, it’s that engineers are now supervising instead of typing. That changes the rhythm of work. If this approach holds up long term, we might look back at manual coding the way we look at writing HTML by hand.

Google’s Deep Think Model Cracks Science Problems

Google has released Gemini 3 Deep Think, a specialized reasoning model designed to solve complex challenges in physics, math, and engineering. It moves beyond standard chat capabilities to handle messy data and problems with no single correct answer.

• The new model scored 84.6% on the ARC-AGI-2 benchmark and 48.4% on "Humanity’s Last Exam," significantly outperforming previous leaders.

• Deep Think uses "test-time compute," meaning it spends more time processing multiple hypotheses before delivering a final answer.

• Early testers have used it to identify logical flaws in peer-reviewed math papers and optimize crystal growth for semiconductors.

• Access is currently limited to Google AI Ultra subscribers, with reports placing the subscription cost around $250 per month.

• The system achieved a gold-medal standard in the 2025 International Math Olympiad and an Elo rating of 3455 on Codeforces.

• Unlike standard models optimized for speed, Deep Think is throttled by a "thinking budget" to ensure accuracy in high-stakes domains.

This release signals a pivot from general-purpose assistants to expensive, specialized research partners for scientists and engineers.

My take: The $250 monthly price tag is the most honest thing Google has done in years. It admits that high-level reasoning is computationally expensive and scarce. We are leaving the era of "free AI for everyone" and entering a reality where access to the smartest synthetic minds costs as much as a car payment each month.

𝐀𝐈 𝐀𝐠𝐞𝐧𝐭𝐬 𝐀𝐫𝐞 𝐄𝐧𝐭𝐞𝐫𝐢𝐧𝐠 𝐏𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧. 𝐈𝐝𝐞𝐧𝐭𝐢𝐭𝐲 𝐈𝐬 𝐭𝐡𝐞 𝐏𝐫𝐨𝐛𝐥𝐞𝐦.

AI agents are no longer experiments.

They’re being deployed into production systems, where they can access data, call tools, and operate continuously without humans watching every step.

But most security systems were built for people and traditional software, not autonomous AI agents.

Teleport just introduced the Agentic Identity Framework to fix this gap.
Check here: https://fandf.co/4a6AOlX

𝐖𝐡𝐲 𝐓𝐡𝐢𝐬 𝐌𝐚𝐭𝐭𝐞𝐫𝐬:
Today, many teams secure AI agents using:

• Static secrets

• Hard-coded credentials

• Legacy IAM tools

This creates credential sprawl, poor visibility, and serious data risk as AI usage scales.

Teleport’s survey of infrastructure leaders confirms it:
69% say AI adoption will require major changes to identity management.

Key Takeaways (Simple Explanation)

1. 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐠𝐞𝐭 𝐫𝐞𝐚𝐥 𝐢𝐝𝐞𝐧𝐭𝐢𝐭𝐢𝐞𝐬: Each agent is treated like a real identity, not a script. This makes it clear who is accessing what.

2. 𝐍𝐨 𝐩𝐞𝐫𝐦𝐚𝐧𝐞𝐧𝐭 𝐬𝐞𝐜𝐫𝐞𝐭𝐬: Access is short-lived and expires automatically, reducing damage if something goes wrong.

3. 𝐙𝐞𝐫𝐨 𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 𝐚𝐜𝐜𝐞𝐬𝐬: Agents only get permissions when needed — nothing is always “on.”

4. 𝐅𝐮𝐥𝐥 𝐯𝐢𝐬𝐢𝐛𝐢𝐥𝐢𝐭𝐲: Every action by an AI agent can be tracked and audited.

5. 𝐎𝐧𝐞 𝐢𝐝𝐞𝐧𝐭𝐢𝐭𝐲 𝐥𝐚𝐲𝐞𝐫: Identity rules are unified across agents, tools, and data instead of being scattered.

𝐏𝐞𝐫𝐬𝐨𝐧𝐚𝐥 𝐓𝐚𝐤𝐞: AI agents are already inside production systems.
Without a proper identity foundation, they introduce uncontrolled risk. Teleport’s framework makes one thing clear: Identity is becoming the trust layer for AI.

Superhero Tools (Productivity)

SciSpace Copilot: This browser extension explains complex text, math, and tables in research papers while you read.

Perplexity: This conversational answer engine provides direct responses to your questions with cited sources for verification.

Scite: This platform tracks how research papers have been cited to show if their findings were supported or contrasted by others.

Paperguide: This research assistant helps you manage references, summarize PDFs, and draft content with accurate citations.

Elicit: This tool finds relevant papers and extracts key data into a table to help you analyze research trends

Quick AI News Bites

• Meta has launched a feature that gives you more agency over your feed. The new tool lets you directly instruct the recommendation engine on what content you actually want to see. This moves curation from passive clicks to active, user-defined choices.

• Elon Musk has unveiled a unified directive for his two major companies. SpaceX and xAI will collaborate to build Moonbase Alpha using autonomous construction bots. This shifts the immediate priority from Mars colonization to a permanent lunar presence.

• OpenAI has trained a major model without relying on NVIDIA hardware. They utilized alternative infrastructure to break the hardware monopoly that currently dominates the market. This development could force a shift in how compute pricing is structured globally.

• The creators of Claude have secured a massive capital injection. Anthropic raised a fresh $30 billion to expand their compute capabilities. This valuation of $380 billion cements their position as a primary challenger in the foundation model race.

• Apple’s roadmap for a smarter Siri has hit a technical speed bump. Internal testing shows that the iOS 26.4 update is too unstable for public release. Engineers are now delaying the major AI overhaul until version 26.5 or later.

Boost revenue and gain new customers by partnering with us

Reach over 35K AI enthusiasts with your product.

Join our newsletter to connect with tech professionals, investors, engineers, managers, and business owners worldwide. DM now!

Got feedback, a story worth toasting, or a wild tech question? I’d love to hear from you; just reply, or find me on X.

Cheers,

— Poonam Soni

X: @CodeByPoonam