ИИ - Data Science / Machine Learning / Big Data

ИИ - Data Science / Machine Learning / Big Data

@datascience_ai · VK

In the catalog since September 18, 2026 · metrics are recalculated every hour

IT & Technology Quality 80/100 Reliability 90/100
Open channel

The channel’s 30-day history (8 messages) will appear in your feed right away.

subscribers
17K
+1 per day
over 7 days
−11
over 30 days
−22
posts over 30 days
8
0 ads
post reach
356
average 363 · ERR 2.12%
ER
0.45%
reactions ÷ reach

About the channel

Только для истинных нейронщиков и интеллектуалов

ИИ - Data Science / Machine Learning / Big Data is a VK community in the “IT & Technology” category. As of November 18, 2026 it has 17,143 subscribers; over 30 days it lost 22. It publishes 0.3 posts a day on average, a typical post reaches 356 (2.1% of subscribers). Engagement rate (ER): 0.4%. Latest post: 2 days ago. TeleMax data, updated every hour.

How the numbers are calculated

ER = reactions ÷ reach, ERR = views ÷ subscribers. Reach is the average post views over 30 days; the median next to it shows a typical post, and unlike the average it does not rise because of a single paid post. Latest post: 2 days ago. The subscriber count was measured 1 hour ago, other metrics were recalculated 1 hour ago.

Subscribers over 30 days

from 17,142 to 17,166 · for the period −22
17 142 17 154 17 166 19.10 25.10 31.10 06.11 18.11
17 142 17 154 17 166 19.10 25.10 31.10 06.11 18.11
DaySubscribersGrowth PostsAdvertising Per day Total Reactions
18.11.2026 17,143 +1 0—— — —
17.11.2026 17,142 −10 0—— — —
15.11.2026 17,152 −2 10— 344 0
09.11.2026 17,154 0 0—— — —
08.11.2026 17,154 −1 20735 735 3
07.11.2026 17,155 −2 0—— — —
06.11.2026 17,157 0 0—— — —
05.11.2026 17,157 +1 0—— — —
04.11.2026 17,156 −2 0—— — —
03.11.2026 17,158 +1 0—— — —
02.11.2026 17,157 −1 0—— — —
01.11.2026 17,158 0 0—— — —
31.10.2026 17,158 −2 0—— — —
30.10.2026 17,160 0 10278 278 2
29.10.2026 17,160 −2 0—— — —
28.10.2026 17,162 −1 0—— — —
27.10.2026 17,163 −3 0—— — —
26.10.2026 17,166 +4 0—— — —
25.10.2026 17,162 −4 20792 830 7
24.10.2026 17,166 0 20715 715 1
23.10.2026 17,166 +1 0—— — —
22.10.2026 17,165 +1 0—— — —
21.10.2026 17,164 −2 0—— — —
20.10.2026 17,166 +1 0—— — —
19.10.2026 17,165 — 0—— — —

The “per day” column shows how many views the day’s posts gained in their first twenty-four hours: this value can be compared across days and channels. Where there is no measurement yet, an estimate marked “≈” is shown, based on the channel’s own reach curve. “Total” is the accumulated counter of the same posts as of today; it keeps growing. Ads and reactions also refer to that day’s posts, not to the channel over the day; on days without posts there is a dash. A dash in subscribers means the number was not measured that day; we do not fill in gaps.

When people read

Average post reach by weekday and hour of publication, 30 days

Not enough posts for the map.

What they publish

Share of posts by type, with average reach and engagement of each

Not enough data.

How views accumulate

Share of final reach N hours after publication

The curve is built from hourly measurements and appears about a week after the channel enters the top tier.

Channel quality

Channel health: is it read, what it posts and how often
80/100
By reach, engagement, ads and posting recency
  • Reach is 2% of subscribers: the audience barely reads the channel

Reach reliability

Can the numbers be trusted: completeness, outliers, jumps
90/100
By data completeness, outliers and subscriber jumps
  • Only 8 posts in 30 days: too small a sample to judge reach

This score is about our data, not about the channel: it also drops where the numbers are simply incomplete. Low reliability with high quality means “the channel is alive, but it is too early to trust its reach”.

Posts

Popular by views over 30 days, latest by time
24.10.2026 00:30 464 views5 reposts Original

Kimi K3.1 Pro Moonshot выпустила самую сильную свою модель: 2.8T параметров всего, 104B активных, мультимодальность. Упор на длинные агентные задачи: игры целиком в Unreal, Unity и Godot, параметрическое 3D в Blender и CAD, финансовые расчёты с падением галлюцинаций на 65%. Контекстное окно 1 миллион токенов По кодингу, агентам и знаниям авторы ставят её вровень с лучшими закрытыми моделями. Эт…

25.10.2026 12:25 463 views2 reactions2 reposts Original

https://www.alphaxiv.org/abs/2610.02185 "Decoding Looped Transformers Better for (Almost) Free" This new paper from Apple treats an early loop from a looped transformer as a rough draft and the final loop as the refined answer. It then uses the difference between them to push the final prediction further in the direction the model was already improving, with no extra training. On Ouro-2.6B-Thinki…

08.11.2026 07:25 406 views1 reactions1 repost Original

GPT-6 Astra can autonomously edit Qwen3.5-0.8B's weights and raise its MATH-500 from 7.9% to 48.1%, without gradients or retraining! "Sparks of Weight-Space Programming" This blog from MIT asks whether LLMs can read and edit neural network weights directly, much like code. In Qwen, Astra traced repetitive reasoning loops to a direction written by two MLP layers, then edited the weights to suppress…

25.10.2026 17:36 367 views5 reactions2 reposts Original

LLM agents can tamper with their own traces that's used to monitor them. This paper shows that popular agents can delete or modify their own logs, and can even discover trace deletion themselves when it improves reward. And their proposed fix is to simply log externally, basically outside the agent’s control. https://www.alphaxiv.org/abs/2609.30266

15.11.2026 19:28 344 views1 repost Original

Interesting work on video assistants that speak up on their own. The bottom line is that answering when asked is the easy part. The hard part is knowing when to say something — and noticing that the user has changed their mind. Most video models today wait for an explicit request. They miss implicit goals, like advice on the next step while someone is clicking through a website, and they don't u…

08.11.2026 07:10 329 views2 reactions2 reposts Original

The Agent Board: Kanban for Machines; Durable Coordination for Agents That Outlive the Conversation https://arxiv.org/pdf/2611.04100v1 Harnesses for large language models have converged on a single-threaded shape: one agent, one history, one machine, one sitting. That shape fails work that spans many agents, outlives any one model call, or waits on systems that are not agents. We describe the Ag…

15.11.2026 19:28 344 views1 repost Original

Interesting work on video assistants that speak up on their own. The bottom line is that answering when asked is the easy part. The hard part is knowing when to say something — and noticing that the user has changed their mind. Most video models today wait for an explicit request. They miss implicit goals, like advice on the next step while someone is clicking through a website, and they don't u…

08.11.2026 07:25 406 views1 reactions1 repost Original

GPT-6 Astra can autonomously edit Qwen3.5-0.8B's weights and raise its MATH-500 from 7.9% to 48.1%, without gradients or retraining! "Sparks of Weight-Space Programming" This blog from MIT asks whether LLMs can read and edit neural network weights directly, much like code. In Qwen, Astra traced repetitive reasoning loops to a direction written by two MLP layers, then edited the weights to suppress…

08.11.2026 07:10 329 views2 reactions2 reposts Original

The Agent Board: Kanban for Machines; Durable Coordination for Agents That Outlive the Conversation https://arxiv.org/pdf/2611.04100v1 Harnesses for large language models have converged on a single-threaded shape: one agent, one history, one machine, one sitting. That shape fails work that spans many agents, outlives any one model call, or waits on systems that are not agents. We describe the Ag…

30.10.2026 20:38 278 views2 reactions Original

Impressive NVIDIA paper on cutting agent costs. For long-horizon agent tasks, you want the capable model for the early steps and a cheaper model for the rest. Handing off in the other direction saves little or nothing. AgentBaton learns when to make that one switch from patterns in task progress. Each check has a median latency of 338 ms on one H100. Across 11 model pairs and two benchmarks, it cu…

25.10.2026 17:36 367 views5 reactions2 reposts Original

LLM agents can tamper with their own traces that's used to monitor them. This paper shows that popular agents can delete or modify their own logs, and can even discover trace deletion themselves when it improves reward. And their proposed fix is to simply log externally, basically outside the agent’s control. https://www.alphaxiv.org/abs/2609.30266

25.10.2026 12:25 463 views2 reactions2 reposts Original

https://www.alphaxiv.org/abs/2610.02185 "Decoding Looped Transformers Better for (Almost) Free" This new paper from Apple treats an early loop from a looped transformer as a rough draft and the final loop as the refined answer. It then uses the difference between them to push the final prediction further in the direction the model was already improving, with no extra training. On Ouro-2.6B-Thinki…

24.10.2026 15:52 251 views1 reactions1 repost Original

Brilliant work on using many harness configurations to scale agentic RL. https://arxiv.org/abs/2610.29458 Many approaches use mainly one fixed harness, which specializes the model but doesn't generalize well. This work uses many harness configurations as a scaling dimension. But the big question is whether it can generalize to harness configurations never seen in training. They pick a compact tr…

24.10.2026 00:30 464 views5 reposts Original

Kimi K3.1 Pro Moonshot выпустила самую сильную свою модель: 2.8T параметров всего, 104B активных, мультимодальность. Упор на длинные агентные задачи: игры целиком в Unreal, Unity и Godot, параметрическое 3D в Blender и CAD, финансовые расчёты с падением галлюцинаций на 65%. Контекстное окно 1 миллион токенов По кодингу, агентам и знаниям авторы ставят её вровень с лучшими закрытыми моделями. Эт…

18.10.2026 14:26 29 views1 reactions Original

98% accuracy. Less than $3 per problem. Research-level theoretical physics. We are witnessing an extraordinary acceleration in AI's scientific capabilities. A fascinating new paper, “TheoPhysEval: Assessing AI Agents at Autonomous Theoretical Physics Problem Solving,” evaluates frontier AI agents on 50 previously unpublished, research-level theoretical physics problems. The results are remarkab…

09.10.2026 09:59 184 views Original

llmgw.ru ии-токены. без vpn. с российских карт

Who writes about the channel

Forwards, reposts, links and mentions in Telegram, MAX, VK, Odnoklassniki and on websites.
How much is written about you
Posts about the channel, distinct sources and their combined audience over 30 days.
Write most often
Top five sources: name, platform, subscribers, how many times they wrote.
Posts about the channel
Every post: who wrote it, where and when, an excerpt, views, reactions and a link.
Mention tone
Negative and positive for each post and their share in the summary.

Similar channels

TeleMax is a catalog and monitoring service for MAX, Telegram, VK, Odnoklassniki and Rutube channels and news websites.

Data is collected automatically from public sources. We are not the authors of channel content and are not responsible for it. Daily Brief headlines and summaries are written by AI from the posts themselves. AI can make mistakes and misstate facts, so check the original sources linked on each story page before using them.

© 2026 TeleMax · 18+