F101: Прямой эфир по пятницам, о жизни программистов

Stream setup & audience

• The host checks that the YouTube stream works (audio/video OK).
• Mentions a long break from streaming and plans to run future sessions only on YouTube (pausing VK).

Summer travel highlight (Kaliningrad / Kaliningrad Oblast)

• Went to Kaliningrad for the first time and enjoyed the Kaliningrad region more than the city itself.
• Especially liked the coastal/tourist area:
  • Clean, high-quality trails
  • Beautiful housing/architecture
  • Green spaces and parks
  • Beaches
• Recommends:
  • Visiting the region (rather than only the city proper)
  • A short flight from Moscow (about 2 hours)
• Mentions coastal/tourist towns:
  • Yantarny
  • Solnechnogorsk
  • Zelenogradsk
• Notes historical German city names and the post-war renaming, calling the renames somewhat confusing/funny.

iPhone “Duo” / Apple criticism (design & product thinking)

• Responds to questions about a newly released “iPhone Duo.”
• Strongly criticizes Apple’s direction:
  • Says it’s a shame/disgrace to make products like that
  • Believes Apple’s past design identity was more coherent and visionary
  • Argues modern products follow marketing/“the boss’s will” rather than true design conviction
• Broader management philosophy:
  • Warns that obeying hierarchical bosses can degrade “product thinking”
  • Suggests real quality comes from designers/architects who can think independently (not just execute specs)

Reading & philosophy detour: Immanuel Kant

• While in Kaliningrad (associated with Immanuel Kant), visited Kant-related places/museum.
• Learned that many “historic” locations (teaching place, apartment, university) are reconstructions.
• Started re-reading Kant and describes frustration:
  • Calls Kant’s writing extremely difficult and often confusing
  • Feels there’s little evidence behind key moral claims
  • Criticizes interpretation-heavy analyses that may miss what Kant actually argued
• Keeps studying anyway, but is dissatisfied overall.

“Wipe coding” / coding with AI: feelings of degradation & how to stay sharp

(Host repeatedly refers to “wipe coding,” meaning AI-assisted coding—likely Claude Code/coding agents.)

Experiences while using AI coding agents

• Host increasingly does not code manually:
  • For the last 5–6 months, claims they wrote no code by hand (only relied on AI)
• Reports “degradation” feelings:
  • Feels like a student behind “child prodigies” (AI peers)
  • Understands some explanations briefly, then loses understanding and falls behind mentally
  • Feels the distance widening as AI outputs improve

Recipe to reduce degradation (explicit advice)

• To avoid degrading, read and understand what the agent outputs.
• Use interaction behaviors:
  • Add clarifying prompts
  • Ask what the agent meant
  • Request more detail and examples
• Warns this can feel like trying to “jump into the last carriage” of a fast-moving train—so the host is still searching for better approaches.

Costs & model availability trend

• Claims top models aren’t necessarily becoming more expensive:
  • Says models are generally getting cheaper, but limits/usage “eat tokens” faster
• Mentions budget behavior:
  • Uses multiple cloud code accounts (around $100 each), switching as quotas end
  • Expects increasing usage intensity even if unit cost drops

How AI dev workflow should work (agent orchestration, restrictions, CI gates)

Multi-agent / restrictive-tools approach

• Describes a workflow with Claude Code:
  1. The first agent drafts changes and makes pull requests.
  2. The host checks quality; if rules are violated, they respond by adding/strengthening restrictions.
  3. A second “restriction” agent updates static analysis/tooling to enforce quality.
  4. Then the first agent must fix until PRs become “green” and compliant.
• Emphasizes that ideal quality emerges through iterative barriers (static checks, analyzers, coverage/style rules, etc.).

Example: banning “magic constants” / single-use constants

• Disagrees with codebases that keep constants “at the top” for readability even if used once.
• Implements/extends a Java static analyzer (mentions “kulis”):
  • Rule: if a constant is used only once, prohibit keeping it as a separate constant.
• Runs the cleanup by instructing Claude Code to update the repo, removing such constants.

“Strict constraints beat guessing”

• Encourages:
  • Using as strict tooling as possible (static analyzers, max rules enabled)
  • Tight rules help agents write better without needing them to “guess” correctness
• Notes it used to be painful to introduce strict tools into an existing team/codebase, but with agents you can now apply large compliant fixes.

What the future workflow needs (beyond terminal text)

• Host is ~99% satisfied with Claude Code, but wants improvements:
  • Current interface is “one-dimensional”: command → response → PRs
  • Wants a more visual “multi-color zones” view (green/yellow/red confidence or status) showing subtasks and reasoning progress
  • Says long single prompts/long textual outputs are sometimes hard to read

“Code review” and manual review

• Argues detailed manual review becomes less necessary:
  • Runs an automation routine that checks PRs daily
  • Rejects/holds merges when uncertain; only a small fraction require human attention
• Important caveat:
  • The system should allow bugfixes/defect elimination while blocking “new features” from untrusted merges
  • Warns PRs meant to “fix” can quietly introduce new functionality

Core engineering belief: morality comes from humans

• Claims the key differentiator won’t be programming-language knowledge or frameworks (agents can handle those).
• Instead, what remains valuable:
  • Fundamental understanding of software engineering quality
  • “IT morality” (architectural principles and what is good/bad in design)
• Example:
  • Agent made many changes confidently, but the host later found architectural principle violations and discarded them
• Advice for becoming that moral/architectural compass:
  • Read fundamental old books on software quality/architecture (he cites his own course on quality metrics)
  • Avoid “quick practical” shallow tech-only materials
  • Emphasizes metrics-based quality as “morality”

Language learning & communicating with agents

• Recommends writing prompts/requests in English:
  • Literature/quality concepts are often in English
  • Agent may “think” in English; translation can help
• If English is weak:
  • Tell the agent to use simpler language and explain more

Technical debt outlook

• Claims technical debt doesn’t catastrophically accumulate because:
  • Agents are used to fix/find/report problems and produce compliant PRs
  • Automated checks are relied upon, and things aren’t accepted blindly
• Counterexample risk:
  • If you instruct agents to do “cosmically smart” redesigns using only the prompt, you can get wrong re-architecture (like turning Tetris into a text editor)
• Personal practice:
  • Avoid manual one-letter edits; let the agent do it and run build/tests
  • Each manual action costs the agent’s attention/time, pushing the host to think more systematically about setup/checkers

AI “drug addiction” and tool comparison

• Claims losing AI would feel worse than losing the internet:
  • Without Claude Code, computer use becomes frustrating/empty
• Mentions trying other agents/tools (including local models), but says Anthropic’s setup is far better
• Prefers orchestration over a single “all-knowing” agent:
  • Many subagents + visible process is more controllable than a black-box single giant agent

Lifestyle/professional market & Russia commentary (work/career)

• Says the labor market is in crisis:
  • Fewer vacancies but lots of applicants (cites an approx statistic: 25 people per vacancy)
• Advises Russian engineers:
  • Don’t leave the employer quickly; especially don’t quit if you can’t find a new role
  • Severance pay timelines are often misleading—“rest and travel” won’t happen
  • Labor Code protections exist, so stay put and endure

Big-picture philosophy & geopolitics (war/chaos)

• Expresses discomfort with global logic:
  • Events increasingly contradict explanations
• Suggests “the real war may not have begun yet,” framing recent crises/wars as preparation/testing
• References Kant-style reliance on reason rather than divinity/conspiracies when interpreting events

Programming language project: a pure object-oriented goal (ELAN)

• Discusses building their own programming language (mentions Elan/ELAN).
• Research question:
  • Can a pure object-oriented language (everything as objects: numbers/strings/lines/etc.) reach performance comparable to C/procedural languages?
• Says the industry believes it’s “impossible,” but their goal is to build compiler/runtime so performance is in the same class.
• Mentions an open repository and welcomes contributors (hundreds of PRs/day).

Notable speakers / places / products / references (end list)

Speakers/People mentioned

• Immanuel Kant
• Robert “Uncle Bob” Martin
• Davis Parnas
• Boris Black (credited with Claude Code, per the host’s claim)
• Vasilishin (Roman Nikadrovich; referenced for global-process view)
• Dmitry (listener name in Q&A)
• Kirill (listener insisted on a question)
• Claude / Claude Code (Anthropic’s agent)
• Bob Martin
• Donald Trump, Zelensky, Putin (political references)

Locations

• Kaliningrad
• Kaliningrad Oblast
• Yantarny
• Solnechnogorsk
• Zelenogradsk
• Kant-related sites/museum in Kaliningrad

Products/Platforms

• YouTube (stream platform)
• VK (paused for next sessions)
• iPhone “Duo” (reported as “new”; host criticizes)
• Claude Code
• Anthropic
• GitHub
• VPN (mentioned for access/streaming)
• IntelliJ IDEA (said they stopped using)
• Cursor (mentioned as a category/tool)
• yaterm / iTerm (terminal tools mentioned; host says they use yaterm)

Technical systems mentioned

• Static analyzers
• Checkstyle
• ESLint
• CI/build/tests
• Pull requests, tickets/issues, master branch discipline

Other

• Mentions “Software Quality Matrix (SQM)”
• Mentions a YouTube course on open-source best practices/development discipline
• Mentions talks/lectures (e.g., “24 lectures”) about software quality metrics