TGStat - аналитика каналов и чатов

Подписывайся, чтобы быть в курсе новостей TGStat.

TGStat Bot

Cтатистика каналов и чатов не выходя из Telegram

SearcheeBot

Ваш гид в мире Telegram-каналов

Статистика Избранное

Dot the AI's

Гео и язык канала: Весь мир, Английский

Категория: Технологии

Accessible artificial intelligence and singularity chronicles.
We research and showcase AI services for all professions.
We write AI-related news. Concise and understandable.

Collaboration: @hello_voic

Связанные каналы | Похожие каналы

Гео и язык канала

Весь мир, Английский

Категория

Технологии

Статистика

Избранное

Фильтр публикаций

Скрывать удаленные

Скрывать репосты

Dot the AI's

10 Jun, 16:00

00:12

Видео недоступно для предпросмотра

Смотреть в Telegram

the first 60 minutes of work are behind us 🏁

14 0 0 2

Dot the AI's

10 Jun, 15:48

00:06

Видео недоступно для предпросмотра

Смотреть в Telegram

While working on characters and physics, already 50 minutes in Ultacode mode.

13 0 0

Dot the AI's

10 Jun, 15:01

Yesterday, I didn't have time to test the new Fable 5 (it's now the flagship model from Anthropic).

I've launched it now, let's test it on a classic, which is timeless.

13 0 0

Dot the AI's

9 Jun, 18:06

Как вы оцениваете свой уровень владения AI?

Опрос

Базовый. Пользуюсь браузерными LLM (ChatGPT / Deepseek / Claude)
Начальный. Пробовал(а) Claude Code / Cowork, но это пока сложновато
Средний. Пользуюсь агентами в IDE, есть GitHub, могу создать и задеплоить простой проект
Нэйтив. Частый вайбкодинг, в IDE использую команды, создаю скиллы, использую MCP, память
Нэйтив+. Умею в оркестрацию, красивую арх-ру, создаю большие проекты, разворачиваю их для команды
Не пользуюсь AI
Напишу свой вариант в комментариях
Посмотреть ответы

8 0 0

Dot the AI's

8 Jun, 17:13

6. DesignMDSupply – provide any site's address – get a ready DESIGN.md with its colors, fonts, indents and components to feed to an AI agent and replicate the style.

#design@TochkiNadAI

6 0 0

Dot the AI's

8 Jun, 17:13

Applying cool design to our projects

Collected links to services and platforms we use in the Camp, when we don't want to go through a long path of trial and error in website/web application design step by step.

A micro-selection of resources where you can copy an entire design system or design individual blocks and give them to your agent:

1. Aura – there's a ton of stuff here: ready-made design systems, skills, templates. You can download entire html or separately copy design tokens.

2. 21stDev – a huge library of frontend elements, you can copy individual blocks, buttons, animations.

3. Mobbin – a large archive of mobile design examples and screens.

4. Getdesign – a growing library of design systems from major brands, the go-to place for design when a client says: Make It Like APPLE.

5. Neuform – full-featured design systems and individual blocks, also copied with one button.

8 0 0

Dot the AI's

5 Jun, 17:18

⬆️ look at this stylish video about his new project made by my friend, artist Denis Semnov

9 0 0

Dot the AI's

5 Jun, 17:17

00:21

Видео недоступно для предпросмотра

Смотреть в Telegram

I trained an AI model on early Russian avant-garde. This LoRA creates images in the style of futuristic illustrations by Kazimir Malevich, Olga Rozanova, Vladimir Mayakovsky, Vladimir Tatlin, David Burliuk, Mikhail Larionov, and others. All these artists began with designing early futuristic books published by my great-grandfather Georgy Kuzmin and Sergey Dolinsky during the period 1910-1914. The model is named “Slop in the Face” in honor of the Russian futurists' manifesto “A Slap in the Face of Public Taste” that they released in 1912. For this model, I vectorized over a hundred illustrations from these books. The LoRA file can be downloaded from the website https://civitai.com/models/2670340/slop-in-the-face

9 0 0

Dot the AI's

4 Jun, 21:12

Is it necessary to be polite with neural networks?

Language models don't get offended by rudeness and don't experience any positive emotions from users' politeness and courtesy. But most people still often add that "please" or "thank you" at the end of their prompt (did you recognise yourself?)

We observe etiquette towards something that, strictly speaking, doesn't exist. Why do we do this and how can politeness become a problem? We spoke with the author of the channel Dots on AI to learn about the features of communication with artificial intelligence.

Do you say "thank you" to chat-bots and neural networks?

❤️ — yes, I always thank them just in case of a machine uprising
🤔 — no, it's just an algorithm

8 0 0

Dot the AI's

4 Jun, 21:12

Do you say "thank you" to your agents?

I’ve reflected on this important topic in the cards of the esteemed Knife.

4 0 0

Dot the AI's

4 Jun, 18:45

Peeked in the chat @vibecod3rs, GPT translated into Russian.

Let's grab the device! 🐒

6 0 0

Dot the AI's

3 Jun, 16:49

6 уроков про Hermes.pdf

By the way, if someone hasn't reached Hermes yet but really wants to –

read this short thread from the personal experience of one researcher.

Yesterday, I just translated his article with observations after 2 months of working with Hermes for you and me.

10 0 0

Dot the AI's

3 Jun, 16:44

01:04

Видео недоступно для предпросмотра

Смотреть в Telegram

Hermes is now available in a desktop version (download here)

If the Telegram mode is a quick and convenient way to keep the agent directly in the chat environment, the desktop version is needed for greater control, to work with it without being blind, see context, manage complex tasks, sandbox, sub-agents, and automation.

Curious, I will test it!

8 0 0

Dot the AI's

2 Jun, 18:05

The spring was rich with new open source releases 😎

Several new models are starting to close the gap with closed LLMs — not only in chats but also in production scenarios.

Among the most notable are new Chinese models:

▶️GLM-5.1
▶️Kimi K2.6
▶️DeepSeek V4 Pro

What's especially interesting:

GLM-5.1 can autonomously perform tasks for up to 8 hours, Kimi K2.6 allows managing an entire squad of subagents, and DeepSeek V4 Pro contains a hybrid attention architecture, which provides a great result for long context.

Open source LLMs are increasingly becoming a part of enterprise infrastructure.

You can connect these and over 40+ models in the Evolution Foundation Models serviceby Cloud.ru:

You gain access to popular open source models, which can easily be adapted to business tasks. The models are ready to use — there's no need to deploy inference or write code, just connect via an OpenAI-compatible API.

👉Go to service👈

8 0 0

Dot the AI's

2 Jun, 17:24

Unexpectedly cool example of instructions for a video model (Seedance)

First, generate a static map of the location, then draw a red line (route), and give this map to Seedance (you can also attach references of individual objects and characters).

Here's the prompt from a video about riding a broomstick (can be used as an example and adapted):

First-person broom-riding POV, cinematic ultra-fast one-take chase. Strictly follow the red flight path in the image. The Golden Snitch stays ahead as the main chase target, but it should not stay fixed in the center. It moves left, right, up, and down during the flight, making the chase feel more alive. Only show the broom handle, gloved hands, and red sleeves. Never show the rider’s face. Chase through the stadium, around the towers, over Black Lake, through the bridge arch, and back to the pitch for the final catch. No red lines, no arrows, no broken broom, no duplicate faces, no jump cuts.

4 0 0

Dot the AI's

30 May, 21:27

00:10

Видео недоступно для предпросмотра

Смотреть в Telegram

Only now have I gotten around to testing Google's new video generation model, Omni

I've run it a little bit, fed it a screenshot of the Kemp website, and detailed the prompt.

Quality, understanding of the prompt, generation speed – superb!

But there are some questions about consistency:

1. Texts. Out of 12 generations, only one produced the same text as on the input screenshot of the site. And that was only for the large text. The small text was messed up in all generations.

2. Characters. No matter how I emphasized the importance of maintaining consistency in the prompt, Omni kept reinterpreting the dog each time.

But overall, it's great, of course. I'll be trying it more.

#videogenerative@TochkiNadAI

7 0 0

Dot the AI's

30 May, 20:49

2. Gave it to GPT Image 2 and Nano Banana 2 (enabled Google search grounding for the latter)
3. Then there were 3 iterations of adjustments for each model specifying the context, after which these two images were produced.

The stereo effect is present in each, but as you can see, visually with the very starting image, one model performed very poorly. There is a clear leader here, although even for it, the stereogram itself delivers a blurred result.

Can you guess which model worked on which? 🤖💬

5 0 0

Dot the AI's

30 May, 20:49

A Magic Eye / SIRDS-style autostereogram. The entire image is filled edge-to-edge with a dense, seamless, uniformly distributed repeating micro-pattern of tiny colorful cartoon dogs, paw prints, bones, and dog collars — all elements must be small, equal in size, and spread with completely uniform density across the whole image. There must be NO visible silhouette, NO dark outline, NO shadow, NO tonal overlay, NO burned-in shape of any dog or bicycle anywhere in the visible image. The pattern must look 100% uniform — no region should be darker, lighter, or different in any way. The hidden 3D object (a dog riding a bicycle) is encoded ONLY through subtle horizontal pixel-shift variations in the repeating pattern, invisible to the naked eye, and only perceivable when the viewer diverges or relaxes their eyes. The background is black. The style is exactly like a 1990s Magic Eye book page. Do not draw any dog or bicycle as a visible element. Do not composite or overlay any shape on top of the pattern.

5 0 0

Dot the AI's

30 May, 20:49

Experiments 🐱

Yesterday, I came across a stereogram (magic eye) in my feed.

In my childhood, I had two such books, and it was my first experience with augmented reality. I think they greatly influenced my career path, and I became interested in testing Nano Banana 2 and GPT Image 2, and checking if they can create such images?

1. Generated a prompt:

6 0 0

Dot the AI's

28 May, 19:55

Just released Claude Opus 4.8 🏋️‍♂️🏋️‍♂️🏋️‍♂️

It is available today, priced the same as Opus 4.7: $5 per 1M input tokens and $25 per 1M output tokens; fast mode — $10 / $50.

The main difference from Opus 4.7 is improved quality in coding, agent tasks, and long work sessions, plus noticeably more reliable behavior and fewer errors in reasoning (well, as always with the release of new models)

Details here.

8 0 0

Показано 20 последних публикаций.

235

подписчиков

Статистика канала