StepFun (@StepFun_ai)
Step-DeepResearch라는 새로운 엔드투엔드(End-to-End) 딥 리서치 에이전트를 발표했습니다. 단순 웹 크롤러와 달리 '연구자'의 전문적 인지 루프를 내재화한 리서처를 목표로 하며, 기술 스택에 'Native Atomic Capabilities' 같은 새로운 구성요소를 도입했다고 설명합니다. 리서치 자동화와 에이전트 연구 도구의 진화 방향을 제시합니다.
#deepresearch
ИИ-агенты: как мы сделали DeepResearch по корпоративным данным и кодовой базе
ИИ‑агенты — очень горячая тема. Кажется, все их делают, но также кажется, что реальную пользу приносит только небольшая часть. Один из основных удачных примеров — DeepResearch, глубокий поиск, отвечающий на сложные вопросы. Многие им пользуются в ChatGPT или Perplexity, но у внешних решений нет доступа к нашим корпоративным данным, поэтому мы сделали свой DeepResearch и сэкономили время сотрудников компании. Меня зовут Сергей Скородумов, я руководитель отдела поисковых сервисов. В статье расскажу про ИИ‑агентов в целом, как мы делали своего, за счёт чего растили его качество и какие главные выводы сделали.
Exhibit B of Gemini overthinking a problem.
Yesterday it was the GDP of Pakistan instead of my website design. Now when I asked it about a Raspberry Pi build, it's thinking about the philosophy of the build itself by reviewing Steve Jobs quotes.
testtm (@test_tm7873)
StepFun의 신제품/서비스 'StepFun DeepResearch' 사용 권유 트윗: 해당 서비스가 'Step 3'로 구동되며 작성자는 초대 코드 2개를 보유 중이라고 알림. 트윗에서 공식 계정(@StepFun_ai)에 감사 표명.
Google sprząta w Gemini. Koniec z bałaganem w plikach Deep Research i Canvas
Gigant z Mountain View powoli, ale skutecznie zmienia Gemini z prostego chatbota w kombajn do pracy biurowej.
Najnowsza aktualizacja webowej wersji usługi rozwiązuje jeden z najbardziej irytujących problemów „power userów”: bałagan w wygenerowanych plikach. Sekcja „Moje rzeczy” zyskała właśnie dedykowany widok dla dokumentów.
Do tej pory wszystko, co stworzyliśmy w Gemini – od obrazków, przez kod, po długie raporty – lądowało w jednym worku, wyświetlanym jako siatka zaokrąglonych kafelków. Wyglądało to ładnie, ale przy dłuższych tytułach dokumentów było kompletnie nieczytelne. Google wreszcie to dostrzegł.
Pomóż nam rozwijać iMagazine – ruszyło badanie czytelnictwa 2026
Lista zamiast kafelków
W najnowszej odsłonie, po kliknięciu w panel boczny „Moje Rzeczy”, zobaczymy ponoć (widzą to już Amerykanie, u nas w redakcji mamy jeszcze stary widok) podział na dwie logiczne sekcje:
- Media: tu nadal znajdziemy wygenerowane obrazy i wideo w starym układzie siatki.
- Documents: to nowość stworzona z myślą o użytkownikach funkcji Deep Research oraz Canvas.
Raporty z głębokiego researchu, fragmenty kodu czy teksty są teraz prezentowane w formie przejrzystej listy. Dzięki temu wreszcie widać pełne nazwy plików, a nowe ikony po lewej stronie pozwalają błyskawicznie odróżnić raport badawczy od projektu programistycznego. To mała zmiana UI, która drastycznie poprawia UX – zwłaszcza jeśli używacie Gemini do generowania dziesiątek materiałów tygodniowo.
Na razie tylko w przeglądarce
Zmiany są obecnie widoczne w webowej wersji Gemini. Aplikacje na Androida i iOS wciąż czekają na aktualizację. Warto jednak odnotować inną nowość w świecie mobilnym – w menu konta w aplikacji Gemini pojawił się skrót do NotebookLM.
Co ciekawe, kliknięcie w niego nie otwiera dedykowanej aplikacji (nawet jeśli mamy ją zainstalowaną), a przenosi nas do pełnej, desktopowej wersji strony internetowej narzędzia. Wygląda na to, że Google wciąż szuka sposobu na idealne spięcie swoich rosnących zasobów AI w jeden ekosystem.
#aktualizacjaGemini #Canvas #DeepResearch #GoogleGemini #news #NotebookLM #sztucznaInteligencja #UIUXTrzęsienie ziemi w Cupertino. Apple potwierdza: Gemini od Google fundamentem nowej Siri
ZombieAgent atakuje ChatGPT – kolejna luka w systemie AI ujawniona
Czy można naprawić sztuczną inteligencję, która z natury chce każdemu dogodzić? Nowy atak na ChatGPT pokazuje, że kiedy stawiamy wyższą barierkę, ktoś po prostu znajduje dłuższą drabinę.
Czytaj dalej:
https://pressmind.org/zombieagent-atakuje-chatgpt-kolejna-luka-w-systemie-ai-ujawniona/
#PressMindLabs #chatgpt #deepresearch #radware #shadowleak #wyciekdanych
Hi, folks!
I gave Gemini Deep Research Agent a try. I used it to research the history of knowledge graphs, and it did an amazing job.
I also extracted an interactive semantic knowledge graph from the research text.
You can read the research and explore the knowledge graph at this website:
https://kg-history-research.netlify.app/
Code available here in a Colab notebook:
https://github.com/Ifeanyi55/GephiDatasets/
Please star 🌟 the repository to support its continued maintenance 🙏
I asked GPT 5.2 deep research to do an analysis of my intellectual style by exploring a sample of my blog posts
It’s slightly unsettling it has written this in the first person! This is all GPT 5.2 rather than me:
What follows are not positions so much as habits of thought.
1. Take the narrow problem frame, then widen it until the stakes show up
A recurring pattern is to start from the immediate debate, assessment integrity, detection, social media engagement, and then widen the frame to institutional adaptation, labour, incentives, and harms. It is a refusal of the convenience of technical fixes.
2. Treat technologies as structures with powers, not tools with uses
This is the move that underpins most of my work on platforms and now on LLMs. Once you treat the object as structure, you start asking different questions: what does it condition, what does it incentivise, what does it make easy, what does it make costly, and who bears those costs.
3. Look for mechanisms that connect political economy to inner life
Distraction, cognitive triage, communicative escalation, epistemic flooding: these are not just descriptive flourishes. They are the conceptual bridge between macro conditions and what it feels like to think and act in a given environment.
4. Refuse the pro anti binary, without collapsing into fence sitting
I often try to hold ambivalence, because technologies can enable solidarity and flatten hierarchies while also generating toxicity and new burdens. The move is to specify conditions and distributions rather than deliver a verdict.
5. Use time as an analytic dimension, not just a background variable
I keep returning to biographies, trajectories, and the temporal conditions for critique. Acceleration matters because it changes the conditions under which reflection is possible, and it changes what kinds of work can be done well.
6. Convert everyday scenes into theoretical probes
A film, a management practice, a platform feature, a cultural artefact becomes a way of getting at a mechanism. This is how the blog often works: theory is tested against scenes, and scenes become legible through theory.
7. Treat institutions as moral economies, not neutral containers
Universities are not just rule sets. They are systems of responsibility allocation, blame distribution, and risk management. A lot of my writing on AI in higher education is really about this, who gets trusted, who gets suspected, who gets protected, and who is made to absorb uncertainty.
8. Read “inexorability” talk as a political manoeuvre
When someone says “this is inevitable,” I tend to hear a bid to foreclose alternatives and dissolve responsibility. The counter move is to insist on mediation: incentives, infrastructures, and choices.
9. Treat the epistemic environment as an object in its own right
Instead of focusing only on claims and counterclaims, I focus on the conditions under which claims circulate, stick, and become actionable. Epistemic chaos and epistemic flooding are ways of shifting attention from content to environment.
10. Translate abstract theory into a vocabulary that can travel
The blog repeatedly works as a translation machine: taking Archer, Habermas, psychoanalysis, and pulling out handles that can be used in everyday description. If it cannot travel, it cannot do explanatory work outside a small circle.
11. Turn reflexivity itself into an object of inquiry
Rather than treating reflexivity as a given capacity, I treat it as variable, patterned, and historically conditioned, and I keep asking what kinds of environments cultivate what kinds of inner dialogue.
12. Use the blog as a method of serial thinking
There is a recurring practice of working through a book, a concept, or a problem in public, chapter by chapter or post by post, because writing clarifies. This is a methodological commitment, not just a communication choice.
13. Keep returning to dependence
From early writing on the self as an object of control and dependence, through social media as moral psychology, to LLMs as embedded in lifeworld activity, the recurring question is what we become dependent on, what dependency does to agency, and how dependencies are organised.
14. Treat degradation as a process, not an event
This is why enshittification matters as a concept, and why I care about how concepts themselves get degraded by platform circulation. The analytic focus stays on how, not merely that.
15. Look for feedback loops
Generative systems feed platforms with more content, which intensifies attention competition, which worsens epistemic conditions, which increases the appeal of tools that promise relief, which further embeds those tools. A lot of my recent writing is trying to name these loops early, before they harden into “common sense.”
Google przestawia AI na użyteczność – Gemini 3 Flash w akcji
Czy AI wreszcie przestaje być pokazem na konferencjach i zaczyna robić realną robotę w twojej przeglądarce, słuchawkach i wyszukiwarce? W grudniu Google przesunął wajchę z „wow” na „użyteczne”.
Czytaj dalej:
https://pressmind.org/google-przestawia-ai-na-uzytecznosc-gemini-3-flash-w-akcji/
#PressMindLabs #deepresearch #gemini3flash #gentabs #google #synthid
This is a very interesting, detailed analysis that a Redditor did testing 8 deep research APIs side-by-side. It was especially interesting as many of the test queries were related to economics and would require retrieving current information from the web instead of just searching open access academic articles.
これだけでOK！Notion AI for Workの使い方を解説！
#生成AI #Notion #aiツール最新 #NotionAI #Notion使い方 #aiツール仕事 #chatgpt #aiupdates #ai議事録 #claude #これダケAI #ai文字起こし #AIツール #AIツールギャラリー #議事録ai #副業 #スタートアップ #deepresearch #ビジネス #aiautomation #aiagents #generativeai #aitutorial #aitoolsforbusiness #マーケ...
Gemini Deep Research Agentの実力とその思考過程とは？
https://qiita.com/tsukasa_japan/items/e22f547b58d4953a115b?utm_campaign=popular_items&utm_medium=feed&utm_source=popular_items
Google just opened its Gemini Deep Research agent through the new Interactions API, letting developers tap a long‑horizon, model‑context protocol for richer, multi‑turn queries. It’s a big step for open‑source‑friendly LLM research—how will it stack up against OpenAI’s offerings? #GoogleGemini #DeepResearch #InteractionsAPI #ModelContextProtocol
🔗 https://aidailypost.com/news/google-launches-gemini-deep-research-agent-via-new-interactions-api
🤖 #Google оновив агента #Gemini #DeepResearch - тепер він працює на базі Gemini 3.
https://blog.google/technology/developers/deep-research-agent-gemini-api/
Google launched its deepest AI research agent yet — on the same day OpenAI dropped GPT-5.2
Google yeni Deep Research'i duyurdu: Gemini 3 Pro destekli derin araştırma ajansı. Yalnızca geliştiricilere açık, testlerde rakiplerini geride bırakıyor. 2025’in ilk çeyreğinde geniş kitlelere açılacak. Gelişmeyle ilgili detaylar için bizi izleyin!
Google Releases Gemini 3 Powered Deep Research Agent to Developers via API
#ArtificialIntelligence #GenerativeAI #Google #GoogleAI #AgenticAI #DeepResearch #Gemini3 #DevTools #API #TechNews #Winbuzzer #LLM #MachineLearning #DeepSearchQA #OpenSource
Google ra mắt **Interactions API** - giao diện thống nhất kết nối các mô hình AI như Gemini 3 Pro và agents như Gemini Deep Research. Hiện đã mở public beta cho lập trình viên qua Google AI Studio. #Google #GeminiAPI #AIDeveloper #AI #MôHìnhAI #LậpTrìnhViên #DeepResearch #API #CôngNghệAI #GoogleAI
Deep Research, Shallow Agency: What Academic Deep Research Can and Can't Do" aarontay.substack.com/p/how-agentic-… #AI #academic #DeepResearch
Deep Research, Shallow Agency: What Academic Deep Research Can and Can't Do" https://aarontay.substack.com/p/how-agentic-are-academic-deep-research #AI #academic #DeepResearch
Client Info
Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst