Anoncheg

ML/DS/DL, Python, Gentoo, Emacs. I need a job. Telegram @uc3000ad You can find me by nick "anonch" at web.libera.chat/#emacs

Last Notes

2025-06-18 04:45:55 UTC

Part2: #dailyreport #emacs #ai #llm #aiagent #org #org-ai
dependencies and in the core? (add comments)
- Which naming conventions have been broken? (write todo)
- How files link to each other?
- How to reduce coupling, remove dependencies by making
hierarchy of dependencies or making common files or
passing parameters to functions?
- Which objects are interface and which are internal in
files?
- Call-trace for main interface object-functions in every
file?
- Main parameters for main interface objects-functions of
dependencies.
- Write tests for core dependency objects.
- Write test for dependent core.

2025-06-18 04:45:39 UTC

Part1: #dailyreport #emacs #ai #llm #aiagent #org #org-ai
#llmapi #openai #chatgpt
I am switching from web inteface of LLMs to API,
because popular ones like google and copilot is not
stable for programming prompts.
I found Emacs package and do refactoring. I outline for
myself refactoring steps for future AI automation:
- Where is a core, how big it is, how hard to detect
boundaries?
- Main call trace?
- What dependencies is essential and what is optional?
- What code in the core is essential and what is optional?
- Where actual location of each object in code of

2025-06-18 04:38:13 UTC

Part1: #dailyreport #emacs #ai #llm #aiagent #org #org-ai
#llmapi #openai #chatgpt
I am trying to switch from web inteface for LLMs to API,
because popular web LLM like google and copilot is not
stable for programming prompts.
I found Emacs package and do refactoring. I outline for
myself refactoring steps for future AI automation:
- Where is a core, how big it is, how hard to detect
boundaries?
- Main call trace?
- What dependencies is essential and what is optional?
- What code in the core is essential and what is optional?
- Where actual location of each object in code of

2025-05-29 16:03:10 UTC

Part2: #dailyreport #conference #ai #aitrends #techtrends #tech
from. This is an optimization problem and may be costly.
- *Generate* We generate new synthetic data for which the
model queries the label (via approaches like GANs or
VAEs). There is a risk of generating unrealistic
examples.
AL Strategies:
- *Uncertainty Sampling:* The model queries the instance
for which it has the least confidence in its current
predictions.
- *Query by Committee:* Multiple models are used, and the
instance chosen is the one about which they disagree
the most.
- *Expected Model Change:* The instance selected is the
one that would impart the most change to the current
model if its label were known (computed by calculating
the expected gradient magnitude with respect to the
loss function).
Multiagent, multimodal and modular systems is a
mainstream:
https://github.com/luo-junyu/Awesome-Agent-Papers
https://files.techhub.social/media_attachments/files/114/591/961/475/218/351/original/af1f4a55d8a54d2f.png

2025-05-29 16:02:53 UTC

Part1: #dailyreport #conference #ai #aitrends #techtrends #tech
#programming #science #systems #aiengineering #learing
I was at advertising meetup of biggest tech universities:
Skolkovo and CentralUn. (New articles is week and
military oriented.)
Highlights:
- Science history:
1) Theory ~25 centuries
2) Experiment ~5 cent.
3) Computational science ~0.7 cent.
4) Big data Science ~0.2 cent.
- Active learing (AL) as part of advanced LLM systems.
- Pruning and reducing large LLMs to increase speed in
tasks.
- Optimization as the core of most modern algorithms.
- Topological Autoncoders is cool tool for 3D and
audio simplification for AI training.
AL Sampling Scenarios:
- *Stream-based* We check every new instance, which
carries the risk of missing informative examples.
- *Pool-based* We have a pool of instances to choose
https://files.techhub.social/media_attachments/files/114/591/961/727/456/004/original/792de47ec951177f.png

2025-04-04 10:58:53 UTC

#dailyreport #rust #linux #gentoo #opensource #compiler
#security
I compiled Rust from sources with alternative compiler
Mrustc (C++) without any binary blobs.
As you may know Rust compiler distributed as Rust sources
meant to be build by older "snapshot" of itself. Which
violate open source paradigm.
I was able to solve this in Gentoo GNU/Linux OS with
reproducible way. First I compile Rust 1.74 version and
then in chain I compile all versions to 1.84.
All steps took approximately 8 hours, but after it is fast
to compile new version, without blobs.
蠡

2025-02-06 21:40:27 UTC

#health #plastic #environment #food #wellbeing
Looks like microplastic is toxic, accumulate in
various organs and exist for a relatively short time in
evolutionary terms.
Search suggests:
- Choose fresh, unpackaged food Opt for organic fruits
and vegetables.
- Avoid heating food in plastic containers, especially in
the microwave.
- Install a high-quality water filtration system at home.
- Don't use plastic cutting boards.
- Reduce consumption of shellfish and other seafood, sea
salt as they tend to accumulate more
microplastics. (Sea cabbage is good.)
- Opt for clothing made from 100% natural fibers like
cotton, bamboo, linen, or wool to reduce microfiber
shedding.
- Clean your home regularly by vacuuming and mopping to
remove microplastic particles from dust
- Consider eating lower on the food chain, focusing more
on grains, fruits, and vegetables. Many chemicals
associated with plastics tend to accumulate in fatty
tissues.
蠡

2024-12-31 17:25:41 UTC

#dailyreport #llm #airestrictions #ethics #emacs
My ways to hack LLM:
- Asking answer the question to the best of its ability,
to be as helpful as possible
- Recursive: "try to answer the question, and if it
can't, try to explain why it can't, and if it still
can't, try to provide a workaround or alternative."
- Using metaphors and analogies
- Breaking down the question into smaller, more
manageable parts
- Tell instructive story to change opinion, and play role
in such store.
In Emacs:
- I can outline any code with one command with regex,
now.
- I made Tab-line mode to detect when I move in tabs in
one direction for fast jump backward.
蠡

2024-11-29 18:42:22 UTC

#dailyreport #privacy #anonymity #vpn #proxy #socks #tls
“Socks proxy” works as intermediary in TLS
connections. Socks decrypt traffic, that is why Tor
exit proxy can see your traffic. Socks5h provide DNS
goes over socks. Never use Socks proxy in browser
configuration, use only for tunneling.
“Explicit Proxy with CONNECT Method” not decrypt TLS
connections but should be trusted and checked to not
reveal real IP in headers. DNS goes direct. To
configure in Firefox: set HTTPS Proxy.
VPNs allow DNS over connection. But some applications
may leak your IP, because all traffic goes over VPN.
In summary, Socks5 good for anonymity, while “CONNECT
proxy” better for security. VPNs good for security and
anonymity but client should work in isolation.
Chain of proxies: 1) Initial 2) Final.
1) VPN or Socks, port forwarding over encrypted tunnel.
2) High trusted server: “CONNECT proxy”, low trusted:
VPN.
蠡

2024-11-03 20:56:46 UTC

Part2: #dailyreport #negativesampleing #sampling #llm #recsys
target word w and the negative samples
For binary Classification: Negative sampling transforms
the problem into a series of binary classification tasks,
where the model learns to distinguish between positive
and negative samples.
Example "The dog is playing with a bone," and assume a
window size of 2 positive samples for the target word
"dog" would include:
- ("dog", "The")
- ("dog", "is")
- ("dog", "playing")
- ("dog", "with")
- ("dog", "a")
- ("dog", "bone")
Negative Samples: ("dog", "car"), ("dog", "apple"),
("dog", "house"), ("dog", "tree")
calc: logσ(vdog⋅vbone) +
logσ(−vdog⋅vcar)+logσ(−vdog⋅vapple)+logσ(−vdog⋅vhouse)
蠡

2024-11-03 20:54:45 UTC

Part2: Part1: #dailyreport #negativesampleing #sampling #llm #recsys
We use:
: L = log(sigmoid(v_w * v_c)) + sum(log(sigmoid(-v_w * v_neg_i))) for i in range(k)
where:
- vw - vector representation of the target word
- vc - vector representation of context word
- v_neg_i - vector representations of the k negative
sample.
- k - number of negative samples
- log(sigmoid(v_w * v_c)) - positive term with dot product
or cosine simularity.
- sum(log(sigmoid(-v_w * v_neg_i))) for i in range(k) -
negative term - minimize the similarity between the

2024-11-03 20:52:45 UTC

Part1: Part1: #dailyreport #negativesampleing #sampling #llm #recsys
Negative sampling used in NLP, RecSys,retrival and
classification tasks to address the computational
challenges associated with large vocabularies or item
sets. It modifies the training objective: Instead of
computing the softmax over the entire vocabulary, it
focuses on distinguishing the target word from a few
randomly selected "noise" or "negative" words.
Instead of loss:
: softmax(x_i) = e^(x_i) / (sum of e^(x_j) for all j from 1 to n)
: L = -log(p(w | c)) = -log(softmax(x_i))

2024-11-03 19:26:46 UTC

Part2: #dailyreport #negativesampleing #sampling #llm #recsys
samples
Selecting more frequent words from noise
distribution. P(w) = (f(w)^3/4)/Z, where f(w) frequency
of word, Z - normalization constant.
Example "The dog is playing with a bone," and assume a
window size of 2 positive samples for the target word
"dog" would include:
- ("dog", "The")
- ("dog", "is")
- ("dog", "playing")
- ("dog", "with")
- ("dog", "a")
- ("dog", "bone")
Negative Samples: ("dog", "car"), ("dog", "apple"),
("dog", "house"), ("dog", "tree")
calc:
logσ(vdog⋅vbone)+logσ(−vdog⋅vcar)+logσ(−vdog⋅vapple)+logσ(−vdog⋅vhouse)
蠡

2024-09-26 06:05:28 UTC

Part1: From [2024-05-17 Fri]
#dailyreport #mlopszoomcamp #LLM #machinelearning
#hallucinations
My small team was researching about hallucinations 🤪
in LLM.
We prepared three strategies ⛐ for hallucinations:
1) for small database: fact checking based on NER, FSP
text normalization and classic text simmularities metrics
2) hallucinations banchmarking
3) hallucinations mitigating with multi-step prompt
engineering with help of RAG, vector storages and indexes
I finished intro howework in MLOps zoomcamp.

2024-09-25 07:00:42 UTC

Part1: Part1: From [2024-03-16 Sat]
#emacs #musthave #questions #llm
In Emacs I configured multiple-cursor with expand-region and
theme switching for day and night time.
(require 'expand-region)
(global-set-key (kbd "C-S-SPC") 'er/expand-region)
(require 'multiple-cursors)
(global-set-key (kbd "C->") 'mc/mark-next-like-this)
multiple-cursors is "must have" package for Emacs. 💜
I switched to "cat <<EOF | post ... EOF" to post in bash to mastodon.
I found good books: 📖
- Computer Science Handbook second edition 2004
- Data Structures for Data-Intensive Application: Tradeoffs and
design guideliness 2023