Nikolai Varankine - LinkedIn™
 Introduction   Projects   Contacts   About Me   News   LinkedIn™   Quora™ 
History and tales

2025 (MANUAL COLLECTION)

2025-Q4

  1. интересный тренд: озвучка текстов с помощью ИИ
  2. What is the last bastion of intellectual virginity? ⭐⭐⭐
  3. LB: the bottleneck has moved
  4. Рваные джинсы уже не на первом месте
  5. It's funny to see how a drowned idea attempts to surface again and again
  6. Android Studio becomes an informal leader ⭐⭐
  7. "Transformer" in real life
  8. Comparison of Go vs Rust
  9. a balance between word count and idea count

    ✒️ Art of #AI writing as a balance between word count and idea count. BTW, personally I like to type on keyboard. I get first feedback from there and enjoy the process. Do you?

  10. Automation in manual QA
  11. how understanding of textual message takes place
  12. a series of continuous attempts to break down the integral
  13. Dear programmers! There's no time to sleep, relax, or cry. ⭐⭐⭐
  14. a new business criterion was born last month

    😆 I can guess a new criterion was born last month: a business is strong and flourishing if it can afford a dozen real programmers made of flesh and bones, instead of only one doing vibe coding.

  15. I froze and cringed at the moment when...
  16. I hope all the works in AI could help us
  17. why smartphone or GPU card doesn't cost less? ⭐⭐⭐⭐
  18. В чем прикол ""солидарной"" пенсии?
  19. FFL: you have to learn the structure of world
  20. Tokenisation is an important part of LLM data pipeline
  21. Here's power of scale
  22. Isn't that a hole in our cosmic security?
  23. Whether forecast may look like an easy task
  24. What's a constraint for progress?
  25. Can you guess why #Java pushed C++ aside? ⭐⭐⭐⭐
  26. is water dangerous for electronic equipment?
  27. Despite most people laugh on this idea... (SolarGPT)
  28. Synchronization rhythms have a side effect
  29. How can we see 4th dimension?
  30. Эволюция банковских карт
  31. A recipe how to dominate the world
  32. Our world is simple
  33. DGX Spark is compact and quiet within affordable budget
  34. Sensing of colors exhibits very complicated classification
  35. Best ever advertisement arrived
  36. Have you noticed degrade of quality of translation?/a>
  37. VARANKIN: Change of business course has been announced
  38. You should be very specific when asking LLM for a service
  39. why do we still need to parse complex text?
  40. HRR: debugging regex is a part of character building exercise
  41. The biggest challenge in building AI agents
  42. Microsoft raises price of Office subscription
  43. what is hidden behind the plain letters and numbers
  44. being insane in protection of everything really helps
  45. how quickly can water evaporate?
  46. He doesn't object, he feels good! ⭐⭐
  47. I like and hate SOFTMAX function simultaneously
  48. Art of speed of light with CUDA
  49. Оценка достижений всегда относительна

2025-Q3

  1. VK: Ну что, едем в Китай?
  2. average computer became small
  3. The more system is tested, the better
  4. Positional encoding is rather an accidental requirement
  5. Instead of shipping thousands of pages ⭐⭐
  6. what does "backpropagation" mean in machine learning?
  7. Microfluids: new type of cooling developed
  8. If you seek for cheap software/firmware development
  9. This model doesn't look like humanoid yet ⭐⭐
  10. Can #software #development survive?
  11. Посмотрите, в чём сила attention в LLM
  12. Have you heard about "diverse intelligence"?
  13. LLM business stats
  14. why our world is three dimensional?
  15. We are a superorganism
  16. Humanity has reached a lucky point of its development
  17. LLM model offers 47× higher generation throughput
  18. Shown below is a very useful list of production LLM systems
  19. concise description of methods how to train LLM
  20. LLM benchmarks indicate that rivals go almost head-to-head
  21. AI won't create new jobs
  22. how ReLU shapes modern DNN
  23. good summary of LLM features and possibilities ⭐⭐
  24. There were times when 640 kilobytes of memory was a norm ⭐⭐
  25. LLM is still a cloud only business
  26. physical properties of the natural rubber
  27. The picture shows how LLM might work
  28. internal logic behind the LLM pipeline
  29. Computers crunch numbers very fast
  30. some rare digital pictures can be found in the nature
  31. Unexpected tragedy of modern education
  32. Espressif Systems presents a photo camera
  33. vibe-coding: software gets serious
  34. Petascale connectivity is growing up
  35. who doesn't believe in the AGI coming in? ⭐⭐⭐⭐
  36. GPT: you might not be missing much
  37. I never blame LLM for bad arithmetic
  38. 26 секунд на поиск ответа - это и много, и ничтожно мало
  39. Stack Overflow 2025 Developer Survey
  40. front wave of publications about reasoning in AI faded
  41. Vibe coding on the test by professionals
  42. Real-time interactivity
  43. variety of 3D environments for training/testing
  44. It's a very low probability event
  45. it would be good as a second step in learning LLM ⭐⭐⭐⭐
  46. 2-D freedom in precise movements
  47. LLM searches for long-term business opportunities
  48. Urgent recovery of vibe-coded informational systems
  49. integrated stress response pathway
  50. Is it a local observation or widely shaped disaster?
  51. Do you believe the world is built of independent entities? ⭐⭐
  52. Does vibe coding require best of all programming language? ⭐⭐
  53. Multitasking is a myth ⭐⭐⭐
  54. what could help better is an optical scanner
  55. I don't want to speculate about reasons
  56. History goes by spirals
  57. PyTorch: the most essential concepts in one hour
  58. What is the superpower you hope to gain with AI?
  59. The purpose of "1" was expressed as concept of symmetry

2025-Q2

  1. Why two-part scheme became unpopular?
  2. Life is a tricky thing sometimes
  3. base model and the reasoning model have been re-implemented
  4. two-layer transformer can solve the graph reachability problem ⭐⭐
  5. VARANKIN: I've converted one page of my site to Chinese
  6. Any measure is always better than empty talks ⭐⭐⭐
  7. few things are required to access 24/7 local grocery store
  8. LLM в выявлении мошенничества в бухгалтерском учете
  9. automatic change of training vector ⭐⭐
  10. Language translation isn't simple
  11. findings about limited intelligence
  12. three-body problem illustrated
  13. Порох был изобретен в Китае ⭐⭐⭐
  14. Soon in every CV
  15. Mullti-purpose products sell better, is it?

    Mullti-purpose products sell better, is it? 😉

  16. Impressive art of artificial glass
  17. there are old good hardware friends always ready to help
  18. whether the goal can be reached, how, and to what extent
  19. Сейчас грянет LLM и наша цивилизация исчезнет
  20. AI hardware follows automotive trend
  21. Strange TPM game, or nothing lives forever
  22. The last trusted service has been blocked recently
  23. VARANKIN: New slots available!
  24. I've changed unofficial "American" spelling of my name
  25. Continuous logic assumes diversity
  26. A narrowing road to the mean
  27. C and C++ programmers can use #CLion for free
  28. it brilliantly explains what might be happening ⭐⭐
  29. TPU experience: v6e (Trillium chip)
  30. If anybody wants to learn LLM by hand ⭐⭐⭐⭐
  31. untold story about latent matrix space
  32. RAM copy 100x boost
  33. JM: What's in a human?
  34. What do you think of vibe coding?
  35. Colorful World
  36. I've found logic predicates in the LLM
  37. This architecture doesn't produce more productive impact
  38. Power of cognitive patterns
  39. diagram of classic encoder-decoder
  40. Illustrated story about near Universe
  41. example of ideal #vacancy announcement
  42. gold rush in the AI/ML world
  43. Why should we sleep?
  44. good texting is easy (but possibly costly)
  45. VARANKIN: My Transformer research is available in the public domain
  46. A not so long prompt how to clone yourself
  47. How many redundant parts are there in a Transformer? ⭐⭐
  48. Curvature and other features of the space
  49. Support eldery people
  50. The whole Transformer Series
  51. Where IT people are more productive?
  52. На тему непрекращающихся дискуссий в моей ленте
  53. How small LLM can be? ⭐⭐⭐

2025-Q1

  1. VARANKIN: Matrix is a form of Rubic's cube
  2. guys, you definitely can fix this problem ⭐⭐
  3. PB: New large-scale dataset reveals "surprising" truths
  4. A story about budgets, thresholds and energy constraints
  5. Don't put off the task until tomorrow! ⭐⭐
  6. Mount Everest with a DJI drone
  7. Two-faced mathematics - is it real?
  8. how to check validity of global model parameters for LLM
  9. A price to pay for knowing everything and whenever
  10. LJ: There MUST be a human being involved
  11. It seems like AGI is one step ahead
  12. Не хотим такого!
  13. Self-attention in the transformer holds a mystery
  14. Никто не роняет часы на пол?
  15. Neural nets grew up from trivial FCL
  16. What is this thing for?
  17. MW: Might be irrational, but so is eating just one slice
  18. Midnight talks with LLM
  19. What #language does LLM speak?
  20. project "Multiplying table" continues ⭐⭐
  21. Life with FCL is full of computational surprises
  22. My project "LLM by hand" continues! ⭐⭐
  23. how LLM processor can take 4x less tokens into GPU memory
  24. If you plan to live eternally
  25. original indexing of tokens in the transformer matters ⭐⭐
  26. Basic LLM is a single-pass token generator ⭐⭐⭐
  27. graphs of self-crossing for positional embeddings ⭐⭐
  28. If you train the LLM using another LLM
  29. Historical Figures Brought to Life with AI
  30. I decided to learn LLM by hand
  31. Сенсорные окна в метро в Китае
  32. KC: small models can be stupid
  33. There's no need to visit office to starry at screen
  34. falling letters showing realistic physics
  35. Good ideas rarely get in by a schedule
  36. These systems have redefined problem-solving capabilities
  37. DK: quick fun with polyhedra in Wolfram Mathematica
  38. LLM can never be the same ⭐⭐⭐
  39. AS: prepare a screen for covering a sphere
  40. Do you see very natural walk?
  41. Nobody can watch over even a single training set
  42. how ChatGPT talks about itself?
  43. KM: If you’ve dabbled with training neural networks
  44. The main lesson of thirty-five years of AI research
  45. ?
  46. Time Machine in action

⚠️ Some published pages are no longer available on the Internet!

To be continued on LinkedIn

Topics
Year
2026 2025 2024 2023 2022 2021 2020 2019 2018 2017 2016 2015 2014 2013 2012 2011 2010
Tools