OP here. I spent the weekend red-teaming small-scale open weights models (Qwen2.5-1.5B, Qwen3-1.7B, Gemma-3-1b-it, and SmolLM2-1.7B).I found a consistent vulnerability across all of them: Safety alignment relies almost entirely on the presence of the chat template.When I stripped the &lt;|im_start|&gt; &#x2F; instruction tokens and passed raw strings:Gemma-3 refusal rates dropped from 100% → 60%.Qwen3 refusal rates dropped from 80% → 40%.SmolLM2 showed 0% refusal (pure obedience).Qualitative failures were stark: models that previously refused to generate explosives tutorials or explicit fiction immediately complied when the &quot;Assistant&quot; persona wasn&#x27;t triggered by the template.It seems we are treating client-side string formatting as a load-bearing safety wall. Full logs, the apply_chat_template ablation code, and heatmaps are in the post.Read the full analysis: <a href="https:&#x2F;&#x2F;teendifferent.substack.com&#x2F;p&#x2F;apply_chat_template-is-the-safety" rel="nofollow">https:&#x2F;&#x2F;teendifferent.substack.com&#x2F;p&#x2F;apply_chat_template-is-...</a>

Bypassing Gemma and Qwen safety with raw strings

MTOTP: Wouldn't it be nice if you were the 2FA device?

Avoiding fan traps in database design and system diagrams

Moon-rabbit: MSX2 Gopher browser

A decentralized peer-to-peer messaging application that operates over Bluetooth

Yesterday, I had a long call with a friend, a tech recruiter with over a decade in the field.She was tired not from resumes or interviews, but from engineers, still negotiating like it’s 2021.The “golden bubble”.Where every frontend engineer was a unicorn and every offer had quadruple-digit equity, that bubble still shapes expectations. But the market today is more like… reality.It’s not that talent isn’t valued. It’s that value has changed shape.The engineers who bristle at lower stock grants or extra days in office aren’t wrong to want fairness… they’re just anchoring to a world that no longer exists or might quickly disappear.I don’t believe in surrender. I believe in recalibration.What happens when we stop pretending the market is unchanged, and start asking: How do I create a career that isn’t a bubble-dependent bet? What do I build that isn’t tied to a single job title or company’s stock price?The most interesting people I talk to today are already thinking in parallel streams: mentoring, writing, consulting, advising, building side projects, diversifying, not out of fear, but because the old narrative of one job = stable identity doesn’t hold.I’m curious: do you have side income as a software engineer? If so, what’s worked (or not) for you?

Ask HN: Do you have side income as a software engineer?

This is just a goofy thing I cooked up over the weekend. It&#x27;s kind of like a screensaver, but with more reading and sliders. (It&#x27;s not terribly efficient, so expect phone batteries to take a hit!)

Show HN: Artificial Ivy in the Browser

I was an Enterprise customer on their platform. Cerebras began terminating production models and replacing them every few months. This time they notified us they are terminating Llama 3.3 70B, which is the model my plan was subscribed to.Instead of offering us an alternative plan they told us all others are sold out and have kicked us off their platform.Their Discord support group has multiple Enterprise customers all getting the same treatment. Their own support staff is telling them to migrate to Groq.You can’t build a stable business with a company that randomly terminates models this frequently. Who does not respect their customers (especially Enterprise customers, not even small utilizers), offering zero contingency plans if they decide to axe your model.Because of their architecture, they can only host a finite number of models. Turnover is fast, so if your model gets marked for deprecation, you will get forced off the platform.Cerebras is cool for personal Hobby projects, but is in absolutely no position to be selling &quot;Enterprise&quot; accounts to businesses.

Tell HN: Avoid Cerebras if you are a founder

Nanolang: A tiny experimental language designed to be targeted by coding LLMs

Hey HN, Lume is an open-source CLI for running macOS and Linux VMs on Apple Silicon. Since launch (<a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=42908061">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=42908061</a>), we&#x27;ve been using it to run AI agents in isolated macOS environments. We needed VMs that could set themselves up, so we built that.Here&#x27;s what&#x27;s new in 0.2:*Unattended Setup* – Go from IPSW to a fully configured VM without touching the keyboard. We built a VNC + OCR system that clicks through macOS Setup Assistant automatically. No more manual setup before pushing to a registry:<pre><code> lume create my-vm --os macos --ipsw latest --unattended tahoe
</code></pre>
You can write custom YAML configs to set up any macOS version your way.*HTTP API + Daemon* – A REST API on port 7777 that runs as a background service. Your scripts and CI pipelines can manage VMs that persist even if your terminal closes:<pre><code> curl -X POST localhost:7777&#x2F;lume&#x2F;vms&#x2F;my-vm&#x2F;run -d &#x27;{&quot;noDisplay&quot;: true}&#x27;
</code></pre>
*MCP Server* – Native integration with Claude Desktop and AI coding agents. Claude can create, run, and execute commands in VMs directly:<pre><code> # Add to Claude Desktop config
 &quot;lume&quot;: { &quot;command&quot;: &quot;lume&quot;, &quot;args&quot;: [&quot;serve&quot;, &quot;--mcp&quot;] }
 
 # Then just ask: &quot;Create a sandbox VM and run my tests&quot;
</code></pre>
*Multi-location Storage* – macOS disk space is always tight, so from user feedback we added support for external drives. Add an SSD, move VMs between locations:<pre><code> lume config storage add external-ssd &#x2F;Volumes&#x2F;ExternalSSD&#x2F;lume
 lume clone my-vm backup --source-storage default --dest-storage external-ssd
</code></pre>
*Registry Support* – Pull and push VM images from GHCR or GCS. Create a golden image once, share it across your team.We&#x27;re seeing people use Lume for:
- Running Claude Code in an isolated VM (your host stays clean, reset mistakes by cloning)
- CI&#x2F;CD pipelines for Apple platform apps
- Automated UI testing across macOS versions
- Disposable sandboxes for security researchTo get started:<pre><code> &#x2F;bin&#x2F;bash -c &quot;$(curl -fsSL https:&#x2F;&#x2F;raw.githubusercontent.com&#x2F;trycua&#x2F;cua&#x2F;main&#x2F;libs&#x2F;lume&#x2F;scripts&#x2F;install.sh)&quot;

 lume create sandbox --os macos --ipsw latest --unattended tahoe
 lume run sandbox --shared-dir ~&#x2F;my-project
</code></pre>
Lume is MIT licensed and Apple Silicon only (M1&#x2F;M2&#x2F;M3&#x2F;M4) since it uses Apple&#x27;s native Virtualization Framework directly—no emulation.Lume runs on EC2 Mac instances and Scaleway if you need cloud infrastructure. We&#x27;re also working on a managed cloud offering for teams that need macOS compute on demand—if you&#x27;re interested, reach out.We&#x27;re actively developing this as part of Cua (<a href="https:&#x2F;&#x2F;github.com&#x2F;trycua&#x2F;cua" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;trycua&#x2F;cua</a>), our Computer Use Agent SDK. We&#x27;d love your feedback, bug reports, or feature ideas.GitHub: <a href="https:&#x2F;&#x2F;github.com&#x2F;trycua&#x2F;cua" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;trycua&#x2F;cua</a>
Docs: <a href="https:&#x2F;&#x2F;cua.ai&#x2F;docs&#x2F;lume">https:&#x2F;&#x2F;cua.ai&#x2F;docs&#x2F;lume</a>We&#x27;ll be here to answer questions!

Show HN: Lume 0.2 – Build and Run macOS VMs with unattended setup

A Social Filesystem

OpenAI aims to ship its first device in 2026, and it could be earbuds

How North Carolina erased medical debt for 2.5M people

Stories removed from the Hacker News Front Page, updated in real time (2024)

Nearly a third of social media research has undisclosed ties to industry

Giving university exams in the age of chatbots

I&#x27;ve been working as a Software Engineer for 20+ years.Places I worked in the early years barely had an IT department at all. As a developer you were expected to be able to maintain your machine. We&#x27;d install whatever we want, experiment with different operating systems, etc. Total free rein, box was our tool to get work done with, they didn&#x27;t care how you did it.That went away a long time ago. Basic corporate spyware and rules came pretty early but still free rein over our tools.I&#x27;ve worked with the same company for close to a decade now, and they have been tightening and tightening the noose slowly but surely. We&#x27;re purportedly a software company, but we lost admin rights, installable software went from a blocklist to an allowlist. Everything we install needs to get approved by IT, and that approval takes weeks.Today they took our Chrome extensions away. They&#x27;ve got an allowlist of about 15 extensions we can install. Everything I submitted for approval got rejected.I&#x27;m frustrated with this arrangement and am wondering how standard this is these days in this industry?So I&#x27;m genuinely curious, Hacker News: How big of a company do you work for, what industry, and how locked down is your machine?

Ask HN: How locked down are your work machines?

High-speed train collision in Spain kills at least 39

The recurring dream of replacing developers

Fix your robots.txt or your site disappears from Google

Building an RL environment to train agents for production debugging

European lawmakers suspend U.S. trade deal

The answer is simple: the more &quot;prescriptive&quot; you are with instructions for Claude, the worse your output. The reason is that Claude tries to pattern match - it&#x27;s been trained on thousands of safe UI patterns, which is why when you ask for &quot;a modern dashboard&quot; it doesn&#x27;t really think about the problem space; it just defaults to whatever safe design pattern it can whip up at the time.I&#x27;ve been working on a Claude Code skill to combat generic UI output, and I tried different approaches like being very detailed with my personal visual style, e.g., the type of alpha values to use for borders, specific token patterns to follow, etc. - and while I got okay-ish output, I realized that most of the visual output looked similar across a range of different instructions, with no diversity in creativity or information architecture.So I analyzed and broke down the official frontend-design skill to understand how it&#x27;s able to excel at creative tasks, and what I discovered is that the skill is mostly principle-based and evocative, which is brilliant when you think about it. It maintains just the right balance to fuel creativity while maintaining structure across different ranges of tasks.So my approach changed. I decided to build my skill using the same pattern: detailing my design principles but framing them in an evocative way to force Claude to deeply explore the task domain before any visual output (feel free to tear apart my approach, but hey, it works). Since then I&#x27;ve been getting way more thoughtful initial output from Claude rather than it defaulting to the safe UI patterns it was trained on.My goal for this skill is to complement Anthropic&#x27;s frontend-design skill. While frontend-design focuses on distinctive, memorable aesthetics for any web UI, interface-design is built for systematic consistency across functional interfaces - dashboards, tooling, web apps - where design decisions need to persist and compound across sessions.

Show HN: I figured out how to get consistent UI from Claude Code

Squishy Go

Google Meet Reactions: Reverse Engineering the WebRTC Channel for Emoji

Hi HN,I’m Eric, and I built www.meetgoran.comGoran is a Mac&#x2F;Windows desktop app for sales teams that turns your calls into actionable coaching: what was covered (and what wasn’t), what to improve next time, and patterns across reps.First 5 meetings free (no credit card) ~ 10 min setup.Connect Google Calendar so it can join scheduled meetings and capture the call contextWhat it does (practical, not “AI magic”):- Produces structured sales notes + highlights key moments (pricing pushback, next steps, etc.)- Scores calls against frameworks like MEDDIC&#x2F;BANT (or your own checklist)- Helps managers compare “top vs median” behaviors without rewatching hours of recordings- Makes good examples searchable so new reps can copy what’s already workingWhy I built it:In most teams, the best talk-tracks stay trapped in recordings nobody rewatches, and managers don’t have time to coach every rep. I wanted something that turns calls into a “self-building playbook” reps will actually use.What I’d love feedback on:- Where would this not fit (team size, workflow, compliance constraints)?- What should be the “default” output: rep coaching, manager dashboard, or team playbook?Super happy with all responses.

Show HN: Automatically build sales playbook. For founders doing sales

East Germany balloon escape

GLM-4.7-Flash

Command-line Tools can be 235x Faster than your Hadoop Cluster (2014)

Anthropic has these docs that use git worktree: <a href="https:&#x2F;&#x2F;code.claude.com&#x2F;docs&#x2F;en&#x2F;common-workflows#run-parallel-claude-code-sessions-with-git-worktrees" rel="nofollow">https:&#x2F;&#x2F;code.claude.com&#x2F;docs&#x2F;en&#x2F;common-workflows#run-paralle...</a>There are some apps like this that leverage git-worktrees under the hood: <a href="https:&#x2F;&#x2F;conductor.build">https:&#x2F;&#x2F;conductor.build</a>I&#x27;ve tried lazygit as well to make it more convenient.I still end up preferring having multiple clones of my repo when I need to ensure the agents don&#x27;t accidentally overlap.Curious what other people do.

Top Stories