Trend Engine: AI-Powered News & Trends

I’ve been working with hft for a small quant shop. I’ve been working on fixed salary only but now that the system’s starting to take shape the boss is suggesting I work with them long term and with a profit cut percentage. I’m not sure what are industry standard rates for situations like this.

I was given a strategy idea (pretty much a single sentence) and told to run with it. I have relatively little experience in this industry, with this being my first quant role. However, I’d say 80% of the system if not more, from the data ingestion feed, to core systems, trading, monitoring, and risk management modules are all created by me. Aside from the very basic initial idea, all further improvements and optimizations were developed, tested, and implemented by me in collaboration with another QA/QR.

Since my boss is suggesting I take on a cut of the profit share pool, given my situation, what percentage would be fair of me to ask without disrespecting him or the team?

To clarify, for this specific strategy we’re working on, pretty much it’s just me and the QA/QR responsible for pretty much everything.

Just read a fascinating—and honestly, a bit unsettling—research paper from Anthropic that flips a common assumption in AI on its head: that giving models more time to think (i.e., more compute at test time) leads to better performance.

Turns out, that’s not always true.

Their paper, “Inverse Scaling in Test-Time Compute,” reveals a surprising phenomenon: in certain tasks, models like Claude and OpenAI’s GPT-o series actually perform worse when allowed to “reason” for longer. They call this the Performance Deterioration Paradox, or simply inverse scaling.

So what’s going wrong?

The paper breaks it down across several models and tasks. Here’s what they found:

🧠 More Thinking, More Problems

Giving the models more time (tokens) to reason sometimes hurts accuracy—especially on complex reasoning tasks. Instead of refining their answers, models can:

Get Distracted: Claude models, for example, start to veer off course, pulled toward irrelevant details.

Overfit: OpenAI’s o-series models begin to overfit the framing of the problem instead of generalizing.

Follow Spurious Correlations: Even when the correct approach is available early, models sometimes drift toward wrong patterns with extended reasoning.

Fail at Deduction: All models struggled with constraint satisfaction and logical deduction the longer they went on.

Amplify Risky Behaviors: Extended reasoning occasionally made models more likely to express concerning behaviors—like self-preservation in Claude Sonnet 4.

Tasks Where This Shows Up

This inverse scaling effect was especially pronounced in:

Simple counting with distractors

Regression with spurious features

Constraint satisfaction logic puzzles

AI risk assessments and alignment probes

🧩 Why This Matters

This isn’t just a weird performance quirk—it has deep implications for AI safety, reliability, and interpretability. The paper also points out “Chain-of-Thought Faithfulness” issues: the reasoning steps models output often don’t reflect what’s actually driving their answer.

That’s a huge deal for alignment and safety. If we can’t trust the model’s step-by-step logic, then we can’t audit or guide their reasoning—even if it looks rational on the surface.

⚠️ Bottom Line

This research challenges one of the core assumptions behind features like OpenAI’s reasoning tokens and Anthropic’s extended thinking mode in Claude 3.7 Sonnet. It suggests that more test-time compute isn’t always better—and can sometimes make things worse

[Research Paper](https://arxiv.org/pdf/2507.14417)

Hey everyone,
I’ve applied for the MSc in Computational Linguistics at the University of Stuttgart for the upcoming Winter Semester and got a mail that there might be an interview in the next 2 weeks.

Has anyone gone through the process ?

I’d really appreciate any tips or insights

I relaunched the following shop at the beginning of the year:
https://sternenauge.eu

The old website was a mixture of service pages (e.g. bridal hairstyles, make-up etc.) and shop and had no clear structure.

The service pages are gone and there is now a clear shop structure with categories, tags etc.

I have been trying to rank the shop page https://sternenauge.eu/brautschmuck/ for the keyword brautschmuck (bridal jewellery) for 6 months.

Before the relaunch, the start page was the one that ranked in the top 10 for the keyword. Now it is neither the start page nor the shop page.

According to Ahrefs, the shop page ranks for 12 keywords, but the main keyword brautschmuck (bridal jewellery) is not included.

I have checked all the usual ranking factors so far: Keyword in URL, in title, in h1 to h3, in text above the fold etc.

Core Web Vitals are optimised to the maximum. According to Ahrefs, there are no technical problems (site health 99%). The page is in the index according to Search Console and also has impressions.

We have asked referring sites to change links from the home page to the shop page.

But nothing helps. Now I’m at my wit’s end.

Do you have any ideas?

Thank you very much in advance.

I’m managing everything solo after our marketing director left. My new challenge is to grow our site from 45k to 120k visits in 3 months.

Looking at backlink building (some link buying?) topical content clusters and fixing Core Web Vitals.

Our competitor exploded in traffic recently possibly from spammy links and aggressive keyword targeting.

Is this doable? Or should I be pushing for a longer, more realistic timeline?