AnswerHN

Hi Hacker News,

I’m thrilled to share Sajoki, a platform we’ve built to revolutionize hiring and talent management using AI-driven personality and soft skill assessments.

The Problem: Traditional hiring processes rely heavily on resumes and subjective interviews, which can lead to inefficiencies, biases, and costly bad hires. Research shows that 89% of hiring failures are due to a lack of soft skills, yet many companies struggle to evaluate these effectively.

What Sajoki Does: Sajoki provides role-based personality tests tailored to specific job requirements. Using AI, we help companies:

Identify the right fit by assessing personality, soft skills, and team compatibility. Reduce hiring costs by up to 40% through smarter, data-driven decisions. Avoid bad hires and build high-performing, complementary teams. Why It Matters: Hiring mistakes can cost businesses up to 30% of an employee’s annual salary, and teams suffer when there’s a poor fit. Sajoki empowers HR professionals to make precise, bias-free decisions while giving candidates a modern, transparent hiring experience.

We’d love your feedback, and feel free to ask me anything about our journey or product.

Looking forward to hearing your thoughts!

Just wonderin

http://www.pennelynn.com/Documents/CUJ/HTML/94HTML/19940045.HTM

I was reviewing byte pair encoding literature and was curious

I'd like some of my online tasks to be automated. What tools are available today to build a custom agent for a website?

Hi everyone,

As the founder of a early stage startup, I’m on a mission to simplify data movement and data quality by building a product that truly meets industry needs. But to create the right product-market fit, I need your help!

If you’re working with data pipelines, ETL/ELT, data quality, or lakehouse technologies, I’d love to hear about the challenges and pain points you face with existing tools.

I’m conducting 45-minute product fit interviews to understand what’s working, what’s not, and what features would make your life easier. If you’re open to sharing your experiences (or know someone who would be), let’s connect!

Your insights will directly influence how we build a flexible, scalable, and user-friendly solution for the modern data stack.

Drop a comment below if you’re interested. Looking forward to the conversation!

Using just scratch.

Why we don't see them used more widely outside of extension cards like graphic cards? Are they much more expensive? Are they slower/have higher latency? Do they consume much more electricity? Or they simply are not the answer to at least part of problems created by multi CPU computing, so no one uses them as "main" RAM. Or there are some other subtle reasons? Do they create new problems for operating system design when used for main computer memory?

I have read over many of the discussions on this site but none of them apply directly to my experiences. My overall problem is that sometimes certain devices which usually operate reliably completely refuse to pair. In all cases trying to forget the device, rebooting the devices in question, shut down, just don't work. 1. I have hearing aids with a BT transmitter for listening to TV. Or rather I have 2 transmitters but only one is flakey. From time to time the 2nd device refuses to pair: the other transmitter is powered down; I reconnect, I reboot, I forget the devices, finally I give up haha. And the next day, unfailingly, everything works as it should. 2. Same thing with my heart rate pickup from a chest belt. 3. Very occasionally I can't reconnect my cell to my car - try all the usual as mentioned above. Connection finally happens some hours later. So all these circumstances happen with equipment used daily, and in respectively unchanging circumstances.

Is it just obtaining a distribution of the next token predictions, or is it more complex?

I came across this linkedin post[0] from: Sr Director, Distinguished Engineer, CTO Office @ Google

The post itself was discussed on HN already: https://news.ycombinator.com/item?id=42852866

> this is my "aha" moment in AI. An AI model that improves other models and itself! with significant proven gains for inference,. Deepseek analyzed the LlaMA code and realized that the KV cache was not optimal so the reasoning engine suggested a way to optimize it. Technically it was probably inspired by Mistral OSS implementation where KV cache uses dynamical allocation. this technique is well know but LlaMA missed it. So, deepseek suggested a code rewrite which was 99% correct with humans only needed to validate and test the code. This improvement results in 2.8x gain for inference on OSS LlaMA. What is even more impressive is that the same change can be adopted to improve DeepSeek itself with initial confirmed gains more that 3x, Effectively a model that is self improving because able to apply previous knowledge when appropriate, augmenting humans and connecting the dots. that's quite something to me. No hype, just pure technological paradigm shift.

Pretty much every sentence has a factual error, I gave DeepSeek the context, the post itself and asked it to fact check:

> The LinkedIn post contains several inaccuracies and exaggerations when compared to the actual news and code provided:

    KV Cache Optimization Claim: The post mentions KV cache optimization as the main improvement, but the provided code and news specifically discuss SIMD optimizations for quantization functions (q4_K, q6_K, etc.) in llama.cpp. There's no mention of KV cache changes in the technical details - this appears to be a conflated claim.

    Performance Gains: While the PR shows 2x speed improvements for WASM through SIMD optimizations, the post inflates this to:

        "2.8x gain for inference on OSS LLaMA"

        "3x gains for DeepSeek itself"
        These specific numbers aren't substantiated in the provided materials.

    Self-Improvement Claim: The post states the model can improve itself, but the evidence shows:

        Humans wrote detailed prompts (multiple iterations needed)

        Humans validated and tested all code

        The model only converted existing ARM NEON code to WASM SIMD

    Mistral Inspiration: The post references Mistral's KV cache implementation, but the actual work:

        Uses Mistral as a reference for model registration patterns

        Focuses on quantization optimizations rather than attention mechanisms

    Code Generation Accuracy: While the post claims "99% correct code", the technical details show:

        Multiple prompt iterations needed (4-8 retries for complex cases)

        Human debugging required for q6_K implementation

        Final code still required human validation

Key Verdict: The LinkedIn post significantly exaggerates and misrepresents the actual technical work. The real achievement was using DeepSeek-R1 to help convert ARM NEON code to WASM SIMD through iterative prompting, resulting in 2x speed improvements for specific quantization functions - not a fundamental architectural improvement to LLaMA or true self-improvement capability.

The post mixes different technical concepts (KV cache optimization vs. SIMD quantization) and inflates the AI's role in the process. While impressive, the actual work is more narrow and human-guided than portrayed.

[0] https://www.linkedin.com/posts/searchguy_this-is-my-aha-moment-in-ai-an-ai-model-activity-7290244226766823425-OUQk?utm_source=share&utm_medium=member_desktop

I have:

Ludicity (https://ludic.mataroa.blog) Steno & PL (https://blog.waleedkhan.name) Mijndert Stujj (https://mijndertstuij.nl) Ben Congdon blog (https://benjamincongdon.me/blog/) ruder. io (ruder.io) Stratechery (https://stratechery.com)

These are basically all ML/AI blogs except Ludicity. Any others I should add?

Hey there, my names Tom Ventura and I'm building a product that I want to integrate with EPICs referall process for psychologists/psychiatrists. However, I need to know more about how that process works in EPIC from a user perspective. I am looking for somebody who will explain to me how the process of referalls works in EPIC and potentially even showing me an empty screen of what you have to fill out if possible, thank you

i know about malware, but not alot about malware and want to add it to my document. I wrote a document about types of malware that are computer specific but now i'm writing another one, what topics should i add? It can be any topic as long as it's related to malware/ and or software that helps with removal and protection against it. Complex topics will take longer to research though since i still have school!:)

A little of numpy in J

  all =: *./
   any =: +./
   sort =: /:~
   shape =: $
   ndim =:  #@:$
   size =: */@:$

   linspace =: 3 : 0
   'a b n' =. y
   h =. (b-a)%(n-1)
   a + h * i. n
   )

   range =: 3 : 0
   l =. #y
   if. l=1 do. res =.  i. y
   elseif. l=2 do. res =.  ({. y) + i. -~/y
   else.  'a b h' =. y
          res =.  a + h * i. >. (b-a) % h
   end.
   res
   )

   sum =: +/
   prod =: */
   cumsum =: +/\
   cumprod =: */\
   diff =: 3 : '2 -~/\ y'
   lcm_reduce =: *./
   gcd_reduce =: +./
   lcm =: *.
   gcd =: +.

   max =: >./
   min =: <./
   argmaxi =: i. >./
   argmini =: i. <./

   argmax =: 4 : 0
   if. x=_ do. argmaxi ,y
   elseif. x=1 do. argmaxi"1 y
   else. argmaxi"1 |: y
   end.
   )

   argmin =: 4 : 0
   if. x=_ do. argmini ,y
   elseif. x=1 do. argmini"1 y
   else. argmini"1 |: y
   end.
   )

   cos =: 0&o.
   sin =: 1&o.
   tan =: 3&o.
   arcsin =: 5&o.
   arccos =: 6&o.
   asin =: arcsin
   acos =: arccos
   atan =: arctan
   hypot =: %:@: +/@: *:
   deg2rad =: (%&180) @: (1p1&*)
   rad2deg =: (*&180) @: (%&1p1)
   radians =: deg2rad
   degrees =: rad2deg
   asinh =: 7 & o.
   acosh =: 8 & o.
   atanh =: 9 & o.
   round =: 4 : '(<.0.5 + y*10^x)%10^x'
   around =: round
   rint =: <.@:(+&0.5)
   floor =: <.
   ceil =: >.
   exp =: 3 : '^y'
   expm1 =: 3 : '(^y) - 1'
   exp2  =: 3 : '2^y'
   log  =:  3 : '^. y'
   log10 =: 3 : '(^. y) % (^. 10)'
   log2 =:  3 : '(^. y) % (^. 2)'
   log1p =: 3 : '^. 1+y'
   add =: +
   reciprocal =: %
   power =: ^
   subtract =: -
   true_divide =: %
   floor_divide =: <.@:%
   float_power =: ^
   fmod =: |~
   mod =: |~
   divmod  =: %,|~
   remainder =: |~
   bitwise_and =: *.
   bitwise_or =: +.
   binary_rep =:   ": @: (2&#.)
   unique =: ~.
   unique_counts =: ~. ,: #/.~
   unique_inverse =: ~. ,: (~. i. ])
   unique_values =: ~.
   unique_all =: ~. ; (I. @: ~:); (~. i. ]) ;   #/.~
   intersect1d =: [ -. ([ -. ])
   setdiff1d =: -.
   setxor1d =: ~. @:(-. , (-.~))
   union1d =: ~. @:,
   shuffle=: 3 :' (?~ #y) { y'

Slide 1 (Title and one-liner) LeetArxiv, Leetcode for implementing Arxiv papers

Slide 2 (Problem) - Programmers are losing their coding jobs to AI. - Becoming an AI researcher is the best path to career longevity. - Understanding math papers is a barrier to becoming a researcher.

Slide 3 (Solution) We teach programmers how to turn intimidating math research into code. By offering annotated, step-by-step, code implementations.

Slide 4 (Market Potential) - 225,000 software engineers were laid off in 2024. - In the same period, AI job postings increased by 2.0%

Slide 5 (Traction, 2 months after launch)

- 586 subscribers on our Substack - 1.22k ARR - 291 customers on Udemy - 505$ ARR

Slide 6 (Business Model) 1. Currently, we offer Substack-based coding guides. 2. We also offer video-based courses on Udemy. 3. We are building a website to host Leetcode-style programming content.

Slide 7 (Team) Solo programmer, Statistics degree

Slide 8 (Ask) -Looking to raise a 20,000 dollar pre-seed for 6 months of runway.

Funds go towards : - hiring a dedicated web developer to maintain the LeetArxiv website. - Cover Docker, MongoDB and Heroku server hosting costs

Here are the slides : https://x.com/murage_kibicho/status/1884626898417864969

Does anyone have a link to a US or EU hosted Deepseek UI and API?

Asking for the obvious reasons - https://theconversation.com/deepseek-why-the-hot-new-chinese-ai-chatbot-has-big-privacy-and-security-problems-248544

and because it is Open Source, it should be able to be lifted and hosted somewhere else than CN.

Hey HN,

As a solo founder, I struggled to keep my LinkedIn personal page and product blog active while running my business. So, I’m building an AI-powered tool that:

Understands your brand & products automatically to create content that feels like you Mixes in real-time industry trends to keep posts relevant Writes, schedules & posts automatically – no manual work needed Learns from engagement data to improve future posts Runs hands-free, so you never have to worry about content again

Would this save you time? I’d appreciate your feedback!

Sort up to some maximum number of items (N items) on a machine with N^2 processors: for each of the N items in the input array, which we we assume are distinct, we will designate N-1 processors to the task of counting in parallel the number of times the item compares greater than another item. In parallel we obtain such counts for every item in the input array. Now we have only to place the N items in their proper order within the output array, again in parallel. Call it what you will, I call it "alreadysort".