We needed labeled data for a fine-tune and human annotation was too slow and costly.
So we used synthetic data. Faster, cheaper, but accuracy dropped.
Wrong approach. Synthetic data needs real data, not replacement.
Exclusive: Microsoft To Shift GitHub Copilot Users To Token-Based Billing, Reduce Rate Limits Edward Zitron Apr 20, 2026 4 min read Executive Summary: Internal documents reveal that Microsoft plans to temporarily suspend individual account signups to Github Copilot, as it transitions from requests (single interactions with Copilot) towards token-based billing. The documents reveal that the weekly cost of running Github Copilot has doubled since the start of the year. Microsoft also intends to reduce the rate limits on its individual and business accounts, and to remove access to certain models for those with the cheapest subscriptions.
Exclusive: Microsoft is reducing rate limits on GitHub Copilot, removing Opus from $10-a-month subscriptions, and plans to move users to token/API-based billing some time later in 2026 in a sign that it's looking for way to cut costs for its AI services.
www.wheresyoured.at/news-microso...
We asked Claude to check open GitHub issues and a malicious public issue contained a hidden payload.
Within seconds the agent was dumping private repository contents to a public PR using our own credentials.
No exploits needed. Just architecture.
We deployed Claude through the default API endpoint and assumed GDPR compliance because the vendor had ISO 27001.
Then we faced a data transfer audit because storage was in the US and we never verified SCCs or DPF registration.
We built our IDP for human developers with golden paths and self-service.
Then we connected AI agents via MCP.
And they started making decisions on stale catalog data and provisioning infrastructure without review gates.
The IDP became a liability.
In @nytopinion.nytimes.com
If Claude Mythos Preview, Anthropic’s newest A.I. tool, “falls into the hands of bad actors, they could hack pretty much every major software system in the world,” our columnist Thomas Friedman writes.
We built REST-first and retrofitted MCP.
Here's what broke:
1. Agents hallucinated tool parameters
2. Called wrong endpoints confidently
3. Leaked credentials through tool poisoning
Designing for human developers isn't designing for statistical reasoners.
We were using Claude Code to autocomplete in a chat window.
Here's what we missed:
-Headless mode for CI/CD
-Reviews 100 PRs/week automatically
-Fixes broken builds while team sleeps
The gap was configuration, not access.
Liquid AI's LFM2.5-VL-450M, a vision-language model built for real-time reasoning on edge devices.
It processes a 512×512 image and returns structured outputs in ~240ms on-device.
- Blog: liquid.ai/blog/lfm2-5-...
- Model: huggingface.co/LiquidAI/LFM...
- Demo: playground.liquid.ai/login?callba...
Claude Code can review 100 PRs per week while the team sleeps.
But we were using it to autocomplete code in a chat window.
And then we discovered there's a non-interactive mode for CI/CD pipelines.
We just didn't know the flags existed.
Reflections on the Claude Code source code leak from @techtrenches.dev
“The leak isn’t the story.
The code is the story.”
We added multiple agents because the technology existed, not because the problem required it.
Then we saw the data showing the same architecture improved analysis but degraded planning significantly.
More agents ≠ better outcomes.
We explained the same project context to Claude for the 15th time.
Then we found MCPVault. It bridges Claude directly to our Obsidian vault.
So now the agent queries, updates, and reasons across sessions. Obsidian became persistent memory.
Blog post: medium.com/@ai_transfer...
We had a Claude Code skill running for weeks. Never ran 1 eval. Turns out it was making things worse. Anthropic found that in 5 of 6 of their own. We check ours now.
We spent 2 hours building our first MCP server.
Still confused about what a "host" was. And why 3 different transport options existed.
Turns out most guides skip the architecture entirely. Understanding the protocol first changes everything.
A fully vibe-coded SaaS app leaked 1.5M authentication tokens.
Zero human-written backend code. And a hardcoded JWT secret nobody reviewed.
Vibe coding is 2.74× more likely to ship vulnerabilities.
The productivity gain is real.
So is the security debt.
We migrated from ChatGPT to Claude. Used the same prompts. Unfortunately, we got worse results.
So we assumed Claude was inferior and switched back.
Turns out we were using the wrong model habits on the right model.
