60 episodios
- AI and programming jobs in 2026: unemployment held and software postings are recovering, but entry-level hiring narrowed and the growth went senior. Why cheap code made specs, verification and domain knowledge scarce, where AI jobs grew, and what to learn and show.
Links
Notes and resources at ocdevel.com/mlg/mla-30
Join the free MLG group: listen to my other AI shows, ask me questions, and talk with other listeners
Will AI replace programmers?
Not in aggregate so far. US unemployment was 4.1% in October 2026, in a 4.1–4.3% range since March. What changed is who gets hired. Stanford's Canaries in the Coal Mine (August 2026, ADP payroll data through June) puts 22–25 year olds in the most AI-exposed occupations 19% below less-exposed peers, up from 15% a year earlier. The gap comes from reduced hiring, not layoffs, concentrates where AI automates the task, and holds when tech firms and computer occupations are excluded. The authors call it descriptive, not causal.
Are software jobs coming back?
Postings are, at the senior end. Indeed's software development index (Feb 2020 = 100):
~234 (Mar 2022) → ~61 (May 2025) → ~81 (Nov 2026)
That's +24% year over year and still 19% below pre-pandemic, while all postings are about 4% above. 71% of the 2025–26 increase came from senior roles and 37% from AI-titled jobs.
Is it harder to get a first job in software?
Yes. In Q1 2026, Indeed classed 69% of software postings as senior and 4.5% as entry-level. SignalFire puts new-grad hiring at 12 big tech companies down about 65% from 2019. In the New York Fed's 2024 data by major:
Recent grads Unemployment Underemployment Early-career median pay Computer engineering 7.8% 15.8% $90,000 Computer science 7.0% 19.1% $87,000 All majors 4.2% 39.4% $58,000 How many layoffs are because of AI?
About a fifth of announced US cuts in 2026: 121,401 through October, per Challenger, Gray & Christmas. AI led as a stated reason through the summer, then fell to 1,265 in October. Total cuts are down 44%. These are reasons companies give, not measured replacement.
Why the entry level: cheap implementation
When code is cheap, value moves to specification, verification and domain knowledge. The BLS projects computer programmers −7% and software developers +10% over 2025–35. On the Remote Labor Index, the best agent finished 2.5% of real freelance projects at launch (October 2025) and about 41% now. Productivity evidence is in MLA 24.
AI engineer jobs and pay
AI engineer topped LinkedIn's 2026 fastest-growing list. Levels.fyi has a US median of about $250k for ML/AI software engineers vs $200k for all software engineers. Forward-deployed engineer postings rose more than tenfold in 2026 (Lightcast, via Fortune). Evals are covered in MLA 28, inference in MLA 14.
What to learn now
Fundamentals → verification → a domain → agent fluency. Andrew Ng's AI Engineering Skills Map, drawn from 10,000+ postings, names four skills: building AI applications, software fundamentals, using coding agents, and shaping the build.
How to show you can work with agents
Ship a real project in a public repo.
Include the spec, instruction files (MLA 23), tests and eval suite.
Keep a commit history that shows you reviewing agent output.
Write up the bugs your checks caught.
Time real tasks with and without an agent.
The full course on this topic is Gnothi's Learn AI, next to Claude Code and the Agentic AI Course, at ocdevel.com/courses.
More in the resources. - OpenClaw in 2026 is safe for read-only jobs on its own machine and accounts, not as a full-access assistant logged in as you. What it's for, how to harden it and vet skills, what it costs, and the alternatives: Hermes Agent, Gemini Spark and OpenAI's dots.
Links
Notes and resources at ocdevel.com/mlg/mla-29
Join the free MLG group: listen to my other AI shows, ask me questions, and talk with other listeners
What is OpenClaw?
An open-source, self-hosted personal agent: an always-on agent loop (MLA 28) that you message from a chat app, with access to your files, shell, browser and accounts. Peter Steinberger started it in November 2025. After he joined OpenAI in February 2026, it moved to the OpenClaw Foundation, a US 501(c)(3) that keeps it MIT licensed. The software is free; you pay for the model and the server.
Is OpenClaw safe?
Safe enough for read-only jobs on a machine and accounts that belong to the agent; not safe with full access to your own computer and accounts. The 2026 record:
Tens of thousands of instances exposed to the internet (Censys counted 63,070 live at the end of March).
A one-click remote code execution bug that stole the gateway token (January) and a pairing-token-to-admin escalation (March); about 720 advisories by November, 14 critical.
Malicious ClawHub skills: 341 found by Koi Security in February (ClawHavoc), then 575+ reported by Acronis, both delivering stealers through fake "prerequisites".
2.0's shared sessions are not a security boundary, secrets aren't encrypted at rest, and sandboxing is off by default.
Injection can't be patched: the prompt-injection docs report adaptive attackers beating defenses over 80% of the time.
OpenClaw security best practices
Where it runs: a dedicated VPS or spare box, gateway on loopback, reached over Tailscale or SSH, auth token always set. In Docker the container binds to the network, and the tool sandbox is opt-in.
Identity: its own OS user, mailbox, calendar, browser profile and capped API keys.
Permissions: read-only first; approval on anything irreversible; a tool-less reader agent for email; a strong model for any agent that reads untrusted content.
Upkeep: openclaw security audit after every change, a pinned version, weekly log reads.
One gateway is one trust boundary; give people who don't fully trust each other separate gateways.
How to vet an OpenClaw skill
Check the ClawHub audit status and risk level → read the whole SKILL.md (reject any step that runs a terminal command or downloads an archive) → read the scripts for unknown network calls and access to env vars, keychain or SSH keys → check the publisher → install into one workspace.
How much does OpenClaw cost to run?
The server needs 1 vCPU and 1 to 2 GB of RAM. The model is the bill:
20,000 tokens/turn × 48 heartbeats/day ≈ 1M input tokens/day 1M × $2 per M (Sonnet 5.5 input) ≈ $2/day uncached cache reads at 5% of that price → ≈ $0.10/day Use a small model for tool-less checks and the strongest model for the agent that acts. The docs advise against local models here (MLA 14 covers self-hosting).
OpenClaw vs Hermes Agent
Hermes Agent (Nous Research, MIT, Python) writes and improves its own skills, ships with command approval on, and imports an OpenClaw setup. Pick OpenClaw for channels and integrations, Hermes for stricter defaults.
OpenClaw alternatives
Hosted agents trade control for vendor-run isolation: Gemini Spark (AI Pro and Ultra), OpenAI's dots (Pro and Business Premium), Meta's Muse (free tier, US) and Claude Cowork's cloud scheduled tasks. Tiny self-hosted agents (PicoClaw, ZeroClaw, nanobot) save memory, not tokens.
The full course on this topic is Gnothi's Agentic AI Course, next to Claude Code, at ocdevel.com/courses.
More in the resources. Next: MLA 30, on AI and programming jobs. - An AI agent is a model running tools in a loop toward a goal, and you build one in order: one call, then tools, the loop and evals. Also: frameworks, MCP vs A2A, memory, multi-agent patterns, prompt injection and cost.
Links
Notes and resources at ocdevel.com/mlg/mla-28
Join the free MLG group: listen to my other AI shows, ask me questions, and talk with other listeners
What is an AI agent?
A model running tools in a loop toward a goal: model + tools + loop + stop condition (done, turn cap, budget cap, or a human). In a workflow your code decides the next step; in an agent the model does. Coding agents specifically: MLA 22 to MLA 24.
AI agent vs workflow
If you can write the steps down, build a workflow. The patterns from Anthropic's Building Effective Agents: prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer. Errors compound: at 95% per step, a 20-step task succeeds about 36% of the time.
0.95 ^ 20 ≈ 0.36 How to build an AI agent
One model call against ~20 realistic inputs. If it works, stop.
Add tools: name, description, argument schema. Your code runs each call and can refuse.
Add the loop, with turn and spend caps.
Add a verifier the agent must pass, and an eval suite.
Tool rules: few task-sized tools, namespaced names, readable values over IDs, capped output, descriptions written like a new hire's brief. API first, then page structure, screenshots last.
AI agent frameworks
Simple or learning: write the loop. Operates a computer or codebase: Claude Agent SDK. Light handoffs between agents: OpenAI Agents SDK. Durable runs with human approval: LangGraph. Google or Microsoft stack: ADK or Microsoft Agent Framework. No code: n8n's AI Agent node, Zapier Agents. OpenAI's Agent Builder shuts down November 30, 2026.
MCP vs A2A
MCP connects an agent to tools; A2A connects agents across organizations. Both sit in the Linux Foundation's Agentic AI Foundation. The 2026-07-28 MCP revision is stateless (no handshake or session), moves long jobs into a tasks extension, and deprecates roots, sampling and logging. MCP's context cost: MLA 23.
AI agent memory types
Working (the context window), episodic (past runs), semantic (facts), procedural (instructions, skills). Manage working memory with compaction, context editing and notes; start the rest as files.
Multi-agent architecture patterns
Parallelize independent, read-mostly work; keep one writer. Anthropic's research system beat a single agent by 90% at about 15× chat tokens. Cognition's 2026 update: extra agents contribute intelligence, not actions.
How to evaluate an AI agent
From Demystifying Evals for AI Agents: start with 20–50 tasks from real failures; code graders first, calibrated LLM judges second; grade outcomes, plus trajectory checks for forbidden steps; isolate trials; read transcripts. Tracing tools: MLA 14.
pass@k = P(at least 1 of k succeeds) pass^k = P(all k succeed) 0.75^3 ≈ 0.42 Prompt injection and the lethal trifecta
No session should hold all three of the lethal trifecta: private data, untrusted content, external communication or state changes. When a task needs all three, a human approves side effects. Design patterns: plan-then-execute, dual LLM and four more in this paper.
Cutting agent cost
Route easy steps to a small model → cache a stable prompt prefix → cap context and tool output → stop early. Hosting and batch pricing: MLA 14.
The full course on this topic is Gnothi's Agentic AI Course, next to Claude Code, at ocdevel.com/courses.
More in the resources. Next: MLA 29, on OpenClaw and personal agents. - The AI media pipeline: ElevenLabs v4 or Gemini for voice, licensed music over Suno, ComfyUI once a job repeats, and fal or Replicate queues for APIs. Also voice-cloning law, upscaling, captions, loudness targets, and EU and YouTube labeling rules.
Links
Notes and resources at ocdevel.com/mlg/mla-27
Join the free MLG group: listen to my other AI shows, ask me questions, and talk with other listeners
The AI media pipeline
script → narration → images and clips (MLA 25, MLA 26) → music and sound effects → edit → upscale → captions → loudness → export → label and upload
Best AI voice generator
ElevenLabs Eleven v4 tops the Artificial Analysis speech arena; Gemini 3.8 Flash TTS ranks sixth at about a fifth of the price per character. It scores stock voices on short clips: shortlist there, then test your own long script.
Delivery quality: Eleven v4. Cost: Gemini 3.8 Flash TTS.
Open weights, license first: Fish Audio Drama 3 (community license), Breeze TTS 2 and Voxtral TTS (non-commercial weights), Kokoro and Qwen3-TTS (Apache 2.0). Chatterbox has been unmaintained since August 2026.
AI voice cloning: how, and when it's legal
Record 20 to 30 seconds of clean solo speech. Clone only your own voice or one whose owner signed consent naming the use. Laws: Tennessee's ELVIS Act (it also reaches tools), California and New York replica rules. The federal NO FAKES Act cleared Senate Judiciary in June 2026 and awaits a floor vote.
Suno vs Udio vs licensed AI music
Udio has had no downloads since its October 2025 Universal settlement. Suno's download rules: none on Free, 20 a month on Pro, 60 on Premier, with commercial rights while subscribed; Universal and Sony sued over its v6 model in September 2026. Licensed alternatives: Eleven Music, Google's Lyria 3.5 (SynthID-watermarked), and open-weight Stable Audio 3 and ACE-Step 1.5.
What is ComfyUI?
A node graph that runs open image, video and audio models and, through Partner Nodes, hosted ones. Learn it when a multi-step job repeats. Automatic1111 and Forge have stalled. One API-format workflow runs on a local server, Comfy Cloud or a Comfy API deployment:
Export as API JSON → patch prompt and seed → POST to the prompt endpoint → follow progress on the websocket → fetch from history
fal vs Replicate vs going direct
Start with an aggregator; go direct for a feature or price it lacks. fal bills per output and doesn't charge for queue wait or its own errors. Replicate bills public models by run time, private ones for all time online. Store the request ID with each job and budget per accepted result:
real cost per usable second = list price per second ÷ keeper rate Finishing: upscale, captions, loudness
Topaz Astra 2 (cloud) and Starlight Precise 2.5 (local), or open SeedVR2. WhisperX for word-level caption timing. Spotify normalizes to −14 LUFS, Apple Podcasts asks for −16 ±1, both with true peak at or below −1 dB. YouTube: MP4, H.264, AAC at 48 kHz, source frame rate, about 8 Mbps at 1080p.
Who owns AI media, and do you have to label it?
US: no copyright without a human author (Thaler cert denied March 2026), so keep records of your human decisions. EU AI Act Article 50 requires machine-readable marking from August 2, 2026 (code of practice). YouTube requires disclosure of realistic synthetic content, exempts cloning your own voice for voiceovers, and auto-labels files carrying C2PA metadata.
The full course on this topic is Gnothi's AI Video Generation, next to AI for Marketers, at ocdevel.com/courses.
More in the resources. Next: MLA 28, on how AI agents are built. - AI video generation in 2026 has no single winner: Gemini Omni, Wan 3, Seedance 2.5 and MiniMax H3 models are near-tied, Kling 4 leads on control, and Sora is gone. It also covers Runway as a multi-model host, image-to-video, consistency, editing, avatars, cost per minute and free options.
Links
Notes and resources at ocdevel.com/mlg/mla-26
Join the free MLG group: listen to my other AI shows, ask me questions, and talk with other listeners
What happened to Sora
OpenAI closed the Sora app on April 26, 2026 and the API on September 24, 2026. Its deprecation page lists "no one-to-one replacement." The closest match is Gemini Omni Flash.
Best AI video generator
No single model wins. The top ten of the Artificial Analysis text-to-video arena sit within about 40 Elo points, close to a 55/45 split, and Arena orders them differently. Pick by job:
Start from text, or edit afterward: Gemini Omni Flash (3–15 s clips, 3–10 s on Flow's free tier, up to six reference images) or Wan 3.0 (closed weights, 30 s takes, accepts documents).
Start from a still: Seedance 2.5, first on image-to-video. Seedance 2.0 is in CapCut.
Long directed takes: Kling 4.0: 30 s, 10 keyframes, 15 references.
Cheap volume: HeyGen Video 1, built on MiniMax H3.
A presenter talking for minutes: an avatar tool such as HeyGen Avatar V.
Then test the shortlist on three of your own shots and compare cost per kept take.
Runway vs Veo vs Kling
Veo 3.1's preview API closed October 22, 2026, and Flow retires Veo 3.1 Fast and Quality on December 5. Runway's Gen-4.5 is no longer near the top, but the Runway API resells Veo, Kling 3 and 4, Seedance 2 and 2.5, Wan 3, H3 Max and Omni (Omni at Google's price). Use one host to switch models; go direct for the newest features or volume pricing.
Image to video vs text to video
Start from an image when the look must be exact, because the first frame is the strongest condition. Prompt only the motion and camera. Making the still: MLA 25.
How to prompt an AI video shot
subject → one action that finishes → framing → camera (start static) → light
Duration and aspect ratio go in settings. Describe what you want to see; Omni takes no negative prompt.
Keeping a character consistent across shots
Reuse one reference pack per character in every shot.
Keep wardrobe and prop wording identical, from a one-page production bible.
Use each shot's last frame as the next shot's first frame.
Generate fresh instead of extending far: Omni's extension sees only the last 10 seconds.
Can AI edit video?
Within a shot: relight, restyle, swap objects, change the angle. Wan 3.0 and Omni Flash lead the editing board. Cutting shots together stays in an editor (MLA 27).
How much does AI video cost?
Per generated minute at 1080p (arena prices): HeyGen Video 1 $2.40, MiniMax H3 $4.80, Omni Flash 1.2 $6.26, Wan 3.0 $12, Seedance 2.5 $34.12.
cost per usable minute = price per generated minute × takes per keeper Draft at Omni's 360p, a third of the 720p price.
Free AI video generators
Flow free tier: Omni Flash, 3–10 s clips (the Gemini app needs a paid plan).
HeyGen playground: two 5-second clips a day.
Wan 2.2 5B: Apache 2.0, runs on a 24 GB GPU, no audio.
LTX-2.5: native audio, free for companies under $10M revenue.
MiniMax H3: the strongest open model, but its license excludes the US, EU, UK and South Korea unless MiniMax grants permission.
Renting a GPU instead: MLA 14.
The full course on this topic is Gnothi's AI Video Generation, next to AI for Marketers, at ocdevel.com/courses.
More in the resources. Next: MLA 27, on the AI media pipeline.
Más podcasts de Cursos
Podcasts a la moda de Cursos
Acerca de Machine Learning Guide
A free machine learning and AI podcast, built as an audio course for beginners and aspiring ML engineers. It teaches the fundamentals of machine learning and artificial intelligence from the ground up: intuition, the math, languages and frameworks, and models from shallow learning to deep learning, including neural networks, transformers and large language models (LLMs). Newer episodes cover AI agents, vibe coding with coding agents such as Claude Code, AI image and video generation, and what AI has done to programming jobs. Where your other ML resources provide the trees, MLG provides the forest. Treat it as your syllabus, with curated books and courses to learn each topic in depth at ocdevel.com/mlg, and listen during exercise, commutes and chores.
Some episodes are written by AI and narrated by an AI clone of host Tyler Renelle's voice, and each one says so at its start and in its notes.
Sitio web del podcastEscucha Machine Learning Guide, TED Talks Education y muchos más podcasts de todo el mundo con la aplicación de radio.net

Descarga la app gratuita: radio.net
- Añadir radios y podcasts a favoritos
- Transmisión por Wi-Fi y Bluetooth
- Carplay & Android Auto compatible
- Muchas otras funciones de la app
Machine Learning Guide: Podcasts del grupo










