What came first: the CNAME or the A record?
27 by linolevan | 7 comments on Hacker News.
Monday, January 19, 2026
New top story on Hacker News: Bypassing Gemma and Qwen safety with raw strings
Bypassing Gemma and Qwen safety with raw strings
12 by teendifferent | 0 comments on Hacker News.
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 <|im_start|> / 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 "Assistant" persona wasn'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: https://ift.tt/8huZiBj...
12 by teendifferent | 0 comments on Hacker News.
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 <|im_start|> / 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 "Assistant" persona wasn'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: https://ift.tt/8huZiBj...
Sunday, January 18, 2026
Saturday, January 17, 2026
Friday, January 16, 2026
New top story on Hacker News: Show HN: Aventos – An experiment in cheap AI SEO
Show HN: Aventos – An experiment in cheap AI SEO
3 by JimsonYang | 0 comments on Hacker News.
Hi HN, we built Aventos- a cheap way to track company mentions in LLMs. Aventos is an experiment we're doing after spending ~6 weeks working on various projects in the AI search / GEO / AEO space. One thing that surprised us is how most tools in this category work. Traditionally, they simulate ChatGPT or Perplexity queries by attempting to reverse engineer the search process. Over the past year, many have shifted to scraping live ChatGPT results instead, since those are signficantly cheaper and reflect more real outputs. Building and maintaining scrapers is tedious and fragile, so recently a number of SaaS products have emerged that effectively wrap a small number of third-party ChatGPT/Perplexity/Google AIO/etc scraping APIs. What felt odd to us is that many of these still tools charge $70–$200+ per month, despite largely being wrappers around the same underlying data providers. So we wanted to test a simple idea: if the core cost is just API usage and commodity infrastructure and software costs are lower because of AI, can we be a successful startup if we price near our costs? What we have so far: 1. Analytics similar to other tools (tracking AI citations, AI search results, and competitor mentions) 2. Content creation features (early and still being improved) We’d love feedback- especially from a non-marketing perspective on: * bugs * confusing terminology or tabs * anything that feels hand-wavy or misleading There’s a demo account available if you want to poke around: username: divit.endal4@gmail.com password: password Happy to answer questions about what other things we've built in the space, how these tools work, etc.
3 by JimsonYang | 0 comments on Hacker News.
Hi HN, we built Aventos- a cheap way to track company mentions in LLMs. Aventos is an experiment we're doing after spending ~6 weeks working on various projects in the AI search / GEO / AEO space. One thing that surprised us is how most tools in this category work. Traditionally, they simulate ChatGPT or Perplexity queries by attempting to reverse engineer the search process. Over the past year, many have shifted to scraping live ChatGPT results instead, since those are signficantly cheaper and reflect more real outputs. Building and maintaining scrapers is tedious and fragile, so recently a number of SaaS products have emerged that effectively wrap a small number of third-party ChatGPT/Perplexity/Google AIO/etc scraping APIs. What felt odd to us is that many of these still tools charge $70–$200+ per month, despite largely being wrappers around the same underlying data providers. So we wanted to test a simple idea: if the core cost is just API usage and commodity infrastructure and software costs are lower because of AI, can we be a successful startup if we price near our costs? What we have so far: 1. Analytics similar to other tools (tracking AI citations, AI search results, and competitor mentions) 2. Content creation features (early and still being improved) We’d love feedback- especially from a non-marketing perspective on: * bugs * confusing terminology or tabs * anything that feels hand-wavy or misleading There’s a demo account available if you want to poke around: username: divit.endal4@gmail.com password: password Happy to answer questions about what other things we've built in the space, how these tools work, etc.
New top story on Hacker News: The Alignment Game
The Alignment Game
4 by dmvaldman | 0 comments on Hacker News.
https://docs.google.com/spreadsheets/d/1BYh9ZtEv4k7xoSXmtf1q...
4 by dmvaldman | 0 comments on Hacker News.
https://docs.google.com/spreadsheets/d/1BYh9ZtEv4k7xoSXmtf1q...