3 месяца запусков c AI: что работает в IT сейчас?

Summary — key technological concepts, product notes, reviews, and recommended practices

Big-picture market dynamics

• Current AI is a heavily funded, partially “bubble” phenomenon driven by large companies (Microsoft, Google, OpenAI, Oracle, Anthropic) and government tenders — a mix of real technological progress and growth/PR tactics (info bubbles, sponsored bloggers, coordinated industry messaging).
• Corporate strategy matters:
  • Big platforms (e.g., Google) can absorb long-term investment and monetize via existing infrastructure (cloud, subscriptions).
  • Smaller AI companies often need faster returns and therefore adopt aggressive pricing and packaging.
• Market effects:
  • Fast releases, high GitHub activity, and frequent “news” generate content for creators but do not always indicate immediate product maturity.

Tools, platforms and product-level observations

• ChatGPT (GPT chat): used as a core assistant for requirements, grooming, and long-context conversations; praised for handling large context reliably.
• Cloud-code / “Claude”-style paid APIs: highlighted as a cost/throughput differentiator (example claim: 10 billion tokens for $200). Cloud-tier offerings can act as loss leaders to acquire users.
• Open-source GitHub projects: very visible and active (large star counts, frequent releases), creating continuous content opportunities and a perception of rapid innovation.
• Cursor (human-in-the-loop/code editor): useful for interactive development and small fixes, but seen as less cost-effective for large-scale token usage compared to cloud-code-priced models.
• Figma and pixel-perfect design workflows: considered increasingly obsolete for early-stage iteration. Prefer rough sketching tools (Excalidraw) or automated AI-generated screens and skip pixel-perfect design until validated by users.
• Notion (with AI/agent integrations) + spreadsheets: recommended for combining conversation, documentation, and lightweight automation (example: Notion + cloud model + SQL/SQLite/vector pipelines for financial agents).
• IDEs: less critical in an agent-driven workflow; much can be orchestrated through scripts, agents, and console tooling. Tests and CI remain important quality gates.
• Local vs cloud inference: local runs and integrations are improving, but cloud services and token bundles can be pragmatic for many workflows.

Agent architecture, workflows and limits

• Core problems: memory management and signal design (how you store, recall, and trigger agent actions).
• Working memory/token limits are real — marketing claims of “millions of tokens” are misleading.
  • Practical working memory likely in the tens to low hundreds of thousands of tokens (speaker cites ~50k–200k and suggests ~150k as comfortable).
• Design recommendations:
  • Keep context minimal and relevant.
  • Orchestrate a small number of parallel agents (speaker limits himself to ~3–4 concurrent agent threads).
  • Use human checkpoints.
• Human-in-the-loop: valuable for requirements gathering and validation; pure agent-driven architecture without a clear human spec tends to be less efficient.
• Workflow stages vary by phase (idea → plan → result) and require different tooling/settings per stage.

Practical product & business advice

• AI speeds information work significantly (example: requirements writing reduced from a week to 4 hours via iterative improvement and tooling).
• Two critical skills to develop: product thinking and systems thinking (these outweigh proficiency in a specific language or IDE). Strong English proficiency is essential.
• Don’t just “go faster” in development; accelerate the right part of the funnel (sales, analytics, segmentation) to convert speed into revenue. Writing code faster alone doesn’t increase income if bottlenecks remain in sales/operations.
• Market selection: prefer solving a specific, high-value problem where asymmetry in access/knowledge exists. Consider B2B (tenders/corporate contracts), B2C, or niche local communities.
• Sales automation: many outreach activities can be automated with agents, reducing the need for traditional cold-calling; large organizations still require corporate-level approaches.
• Risk & strategy: corporations are safer in turbulent times; independent founders should pick a clear problem and market and be prepared to iterate and human-check agent outputs.

Warnings and quality control

• Models still hallucinate: the speaker reports agents lie roughly 16% of the time — always verify outputs, check links and facts before trusting or shipping.
• Over-acceleration risk: speeding up the wrong processes amplifies errors and noise; focus on the highest-leverage targets for improvement.
• Token and cost marketing are often misleading — evaluate real token costs and performance tradeoffs.

Toolset the speaker actually uses (practical stack)

• ChatGPT: large-context conversation and specification writing.
• Notion: documents, specs, simple spreadsheets and integrated agents.
• Excalidraw: sketching; preference for messy/fast visual thinking over pixel-perfect Figma early on.
• Cloud-code (Claude-type): cost-effective for token-heavy model usage and agents; recommended over purely local or cursor-only approaches for many tasks.
• Cursor: interactive coding and human-in-the-loop edits (still used but seen as fading versus cloud-code economics).
• GitHub: code/versioning.
• SQLite + vector stores: data/agent memory examples.

Guides, tutorials, and community offerings

• The speaker plans deeper, paid/closed-community tutorials rather than full YouTube coverage:
  • A series of four lessons (approx. 8 hours total) covering agent workflows end-to-end — memory, signals, building sub-agents, demos and cases.
  • Weekly streams and practical masterclasses in the community focusing on demos and hands-on practice (not just theory).
• The speaker suggests YouTube is not a suitable format for detailed agent workflow training and invites viewers to join the community for practical training.

Concrete examples and anecdotes

• GitHub projects with massive activity/releases are used as PR fodder, inflating perceptions of transformational change.
• Example — the speaker’s uncle (finance professional): used Notion + cloud model + a small pipeline (vector store/SQLite + GitHub commits) to build a working agent that automates domain-specific analysis with minimal coding background.
• EPAM example: corporate investment in employee training and automated flows showed dramatic throughput improvements; some teams later spun off into consulting.

Final actionable recommendations

• Learn system thinking and product thinking; get comfortable architecting agent workflows and handling memory/signals.
• Use ChatGPT + a cost-effective cloud token provider for heavy-context, agent-driven workflows; use Notion/Excalidraw for docs and rapid sketches.
• Focus acceleration on bottlenecks that produce revenue (sales, analytics, segmentation), not just development speed.
• Start with small, focused niches or join a corporation if you prefer stability; be prepared to validate and human-check agent outputs.

Main speakers and sources referenced

• Primary speaker: the channel host (monologue-style), occasionally annotated as “M.” / “Op.” in the transcript but effectively a single presenter.
• Companies/platforms referenced: Microsoft, OpenAI, Oracle, Anthropic, Google, GitHub, EPAM.
• Tools/technologies referenced: ChatGPT (GPT chat), cloud-code / Claude-like APIs, Cursor, Notion, Excalidraw, Figma, Miro, GitHub, SQLite, vector stores, various IDEs/editors (Visual Studio, WebStorm), local vs cloud model deployments.