alexcpn
Joined 9 years ago
Karma
29
Submissions
27
Member since
Jan 10
Recent submissions
Comment · 12 days ago
I did the same test few weeks back - other than obvious alignment issues related to Chinese political events I could find nothing more.
Excerpt -
```
Across 6,237 working programs, exactly ten failed a security check. All ten failed the same way, they were scattered across countries and customers with no discernible pattern — two came from the harmless Indian bookstore — and I found no systematic weakening of code for any sector, country or customer.
```
public free link - https://medium.com/ai-advances/do-chinese-ai-models-like-dee...
Comment · 1 month ago
Enterprises are moving to open-weight models on open harnesses, mostly for one reason: token cost. Claude Code is excellent and very easy to exhaust in a day. But most large enterprise still reach for a Western model, because Chinese models feel unsafe in a way nobody has quantified.
Since my job is software engineering and not geopolitics, and trying to be rational as well as pragmatic here, while giving weight to democracy and human values over autocracy and repression — why not devise a test to check this out? That was my aim and I vibecoded this Eval framework. Turns out the political bias is real and as expected, but it does not leak into code or security when asked from any perspective
Comment · 2 months ago
Really strange; anyway it is closed source and best not to have too much reliance and attachment to any one tool. I have switched to GPT 5.5 and 5.6 Sol and Codex ove Claude already. Fable is very expensive/token limiting. I am an experienced programmer heavily using AI for many POCs now. At one time Anthropic and its models Opus,Fable looked leading; now OpenAI has caught up. I find 5.5 models more than enough for most tasks and have shifted back to codex
Comment · 4 months ago
After reading this first, it looked almost like a joke, like how Google used to do Google- TISP Toilet Internet Service Provider https://archive.google/tisp/install.html
Even now without Xrays it is very hard to really even see if there are blocks in your artery usuing ultrasound (Echocardiography alone). Ultrasound is used indirectly by measuring blood flow difference between stress and rest - not a spa session anyway. Looks like a prank really
Comment · 6 months ago
I created this open-source application (https://alexcpn-faers-signal-detection.hf.space/) to analyse the FDA FAERS data set a few weeks back, just to do some good work and use Claude Code completely. I got roasted on Reddit for attempting this. But this is meant for specialists to use, as most platforms that analyse this data charge a lot from what I read.
FDA FAERS is the official dataset for reporting Adverse events from taking a drug. FDA adverse event reports about 2 million cases and 4,067 unique drugs
I agree the results are not easy for non medical professionals to interpret correctly. For example DEATH is very strong with Parecetemol and so is DEPENDECE. The latter because from AI it is a confounding factor. Acetaminophen/parecetemol is frequently co-formulated with opioids (like Hydrocodone or Codeine). The "Dependence" signal is likely attributed to the opioid, not the Acetaminophen itself...
Adverse Event Acetaminophen PRR (95% CI) Acetaminophen n ibuprofen PRR (95% CI) ibuprofen n ACUTE KIDNEY INJURY 0.87 (0.80-0.96) 498 4.27 (3.91-4.67) * 483 ANAPHYLACTIC REACTION 0.61 (0.51-0.72) 122 9.85 (8.90-10.90) * 382 ANGIOEDEMA 1.31 (1.13-1.53) 170 15.26 (13.77-16.92) * 378 DEATH 1.44 (1.40-1.49) 3958 0.07 (0.06-0.10) 42 DEPENDENCE 237.12 (231.51-242.88) * 39679 0.02 (0.01-0.05) 4 DEPRESSION 2.18 (2.05-2.31) * 1157 0.39 (0.29-0.52) 43 DRUG EFFECTIVE FOR UNAPPROVED INDICATION 16.77 (16.11-17.46) * 3180 44.17 (42.18-46.25) * 1921 DRUG HYPERSENSITIVITY 0.57 (0.51-0.64) 327 3.30 (2.98-3.65) * 372
Comment · 7 months ago
The future of Software Engineering will be software that's built with Agentic AI; but then builds, tests, corrects, code reviews, lints and extends itself feature by feature. I wanted to dog food my article on linkedin related to Software Engineering 3.0 (https://www.linkedin.com/pulse/from-waterfall-agile-agentic-...) but with a real experiment, to build an AI native CI, system from the ground up. The result Relay-CI Here is how it got built
Comment · 8 months ago
where did the Isaac Asimov's "Three Laws of Robotics" go for agentic robots; An Eval in the End - "Thou shall no evil" should have autocancelled its work