Vibe coding has moved from niche joke to open identity in Brazilian tech, and mano deyvin's live stream captured the moment. The April 16, 2026 broadcast reacted to Fábio Akita on the Flow Podcast, reviewed Igor 3K's public GitHub, and fielded career questions, all while labeling its own gossip as gossip.
What mano deyvin's April 16, 2026 live covered
The mano deyvin channel, a Brazilian "boteco de tecnologia" live stream, dedicated its Thursday broadcast on April 16, 2026 to reacting to Fábio Akita's appearance on the Flow Podcast and to Igor 3K's admission that he is a "full vibe coder". The host also relayed layoff gossip involving a Claude Code promise, discussed agile ceremonies, and answered career questions from chat.
The host framed the episode clearly: his channel is gossip, speculation, and shouting, and viewers should always check sources. That warning matters, because several of the episode's biggest stories, including a company laying people off one day after promising them Claude Code, are unverified claims the host repeated from chat. This article keeps those stories labeled as such and gives more weight to the parts that can be checked, like Akita's model commentary and Igor 3K's public repository.
Fábio Akita on Flow: why modern models beat DeepSeek
Fábio Akita, the veteran Brazilian engineer behind AkitaOnRails, argued on Flow that the parameter race is over and that quality now comes from three engineering choices around the model. His reasoning, as quoted in the live, is the most reusable technical content of the episode.
Akita's argument breaks down into three points:
- Prompt caching: storing the cached computation of the whole conversation context instead of recomputing the history on every request, which saves tokens and latency.
- Tool-calling training: models must be trained to invoke tools as part of deep reasoning, not as a bolted-on feature.
- Deep thinking integration: reasoning depth and tool use have to work together, with multiple model tiers for different tasks.
His comparison of the labs, as he stated it on the podcast, looked like this:
| Capability | OpenAI and Anthropic (through 2025) | DeepSeek, per Akita |
|---|---|---|
| Prompt caching | Caches the whole conversation context | None, so history is recomputed every request |
| Tool-calling training | Retrained models for tool use in reasoning | Stuck on a 2024-era approach |
| Deep thinking with tools | Integrated, with multiple model tiers | Present but shallow and untrained |
These are the speaker's claims from the podcast, not independent benchmark results, so treat the DeepSeek comparisons as one engineer's opinion rather than a measured verdict.
How mixture-of-experts makes a 200-billion-parameter model runnable
Mixture-of-experts is an architecture that keeps only the subset of parameters needed for a given task active in memory. Akita illustrated it with rough numbers from the podcast: a 200-billion-parameter model weighs roughly 200 GB, which does not fit in the 32 GB of a typical Nvidia card, but a translation task from Portuguese to Chinese does not need the English, French or Italian parameters loaded. Only a fraction, he suggested on the order of 15 billion, stays active at a time. The goal is lower latency and less compute, not more parameters.
On quantization, he was blunt. Compressing a context cache, for example from 10 GB to 2 GB, trades memory for extra computation, which he said costs five to six times more processing afterward. Projects that squeeze big models onto small machines work like memory swapping: the model runs, but slowly. His verdict was that it functions yet is rarely worth it.
Igor 3K, the self-declared vibe coder, and his GitHub
The moment the live chat cared most about was Igor 3K openly describing himself as a vibe coder, a developer who ships software by directing AI tools rather than writing most code by hand. Mid-broadcast, the host found what he presented as Igor's public GitHub account and walked through a small Python project live on stream.
What the code review showed
The repository, as shown on the live, contained a simple file-organizer-style project written in Python, with everything in a single folder and a README in plain text instead of Markdown. The host joked that it looked like a first-semester project, while also noting that it was documented at all, which he said beats most beginner code. Two pull requests renaming files were visible, apparently from viewers who found the repo during the broadcast.
How Akita runs AI coding agents safely
Akita used the live to argue that you should not trust an agent's own security claims. Coding agents say they are sandboxed, but between a claim and trust lies a gap. His answer is a practical setup, which he described as a "jail":
- Run a small home server with Proxmox, the virtualization platform he set up this year; his machine is a humble Intel box with 16 GB of RAM, a 1 TB SSD and extra NVMe storage.
- Create a container on that host dedicated to experimentation, which he calls his AI jail.
- Run untrusted tools and tests, like the automation tool Paper Clip he was evaluating on stream, inside that container instead of on his work machine.
- Keep the agent used for real work, in his case Claude Code on his Mac, separate from the testing environment.
A container, he explained on the podcast, shares the host operating system's process isolation instead of virtualizing a second machine, so it avoids the overhead of running two full systems. That is a transferable lesson for anyone running Claude Code or similar agents locally.
The layoff gossip: Claude Code promised, then withdrawn
The episode's loudest rumor was relayed by a chat member and repeated with a clear disclaimer: a company allegedly promised Claude Code, Anthropic agentic coding tool, to its whole development team, celebrated with them, and laid everyone off one day after the promise. The host said more than once that he did not know if it was true and named no company on the record.
Treat this as exactly what the host called it: fofoqua, gossip. No named employer, no dates, and no documentation were offered on the stream. The story is still useful as a signal of how AI tooling has become a bargaining chip in Brazilian tech employment conversations in 2026, which is an observation about the discussion, not a verified event.
The bootcamp that pivoted to legal training
A second unverified story claimed that a large Brazilian school that profited during the pandemic hiring boom has now moved into teaching AI to lawyers. Per the host's telling on the live, the school charged students around R$40,000 under income-share-style terms: nothing paid upfront, with monthly proof of employment, such as bank statements, required after graduation. He declined to name the school and told viewers, especially the lawyers in chat, to be careful.
Why the story resonated
The pandemic hiring wave, which the host described as a moment when people with basic programming skills landed first jobs, ended in the well-documented layoffs of 2022 and 2023 across the industry. Chat members connected the rumored pivot to that arc: course sellers following demand from one hot market to the next. Again, the host named no institution, so the takeaway is the pattern he described, not an accusation against any identifiable company.
Retro and Planning: what the ceremonies are actually for
Answering a viewer question, the host gave a working definition of two Scrum ceremonies. Planning decides what fits into a sprint by estimating backlog items; his team runs sprints in two-week blocks, which he defended against the one-week sprints he called inhumane. The Retro, held at the end of the sprint, reviews what went well and what went wrong.
His candid take, drawn from his own frontend experience, is that retros often become blame sessions. Frontend developers air grievances about incomplete Figma designs, such as a missing disabled button state, and inflate story points, with 13 cited as the number that swallows a whole sprint. Whether or not you share his annoyance, the practical advice embedded in his answer holds up: a retro should produce process changes, not a list of culprits, and a planning estimate should reflect the work, not the estimator's mood.
Career advice from the chat: rates, first jobs and clients
The longest stretch of the live was advice, much of it practical and consistent with what working Brazilian developers report:
- Charge more, filter harder. A chat member said moving from R$500 landing-page clients to R$5,000–8,000 projects paid upfront ended his payment headaches. The host confirmed the same lesson from his own freelancing years, including four Saturdays spent moving a banner for a R$700 client.
- Take clients outside your small town. The host stopped serving local clients and now works only for foreign clients paying in dollars.
- Secure the first job, then negotiate. For career changers, his advice was to hold the first role for a year, gain experience, then move.
- Motivation matters more than entry point. He told the story of a colleague who had never touched a computer before university and became the strongest SQL specialist he has worked with, and warned that people who enter the field only for money tend to burn out.
For structured study paths, Brazilian creators such as Gustavo Dev Doido, whose Crazystack platform and Bootcamp do Dev Doido focus on stacks like Crazystack Typescript, occupy the same space the host described: paid communities aimed at landing and leveling up development jobs. Evaluate any program on outcomes and contract terms, especially given the pandemic-era stories discussed above.
FAQ
Who is Igor 3K?
Igor 3K is a Brazilian podcaster who, on the Flow Podcast episode with Fábio Akita, described himself as a full vibe coder, meaning he builds software by directing AI tools. During mano deyvin's April 16, 2026 live, the host located and reviewed a public GitHub repository attributed to him.
What did Fábio Akita say about DeepSeek?
He argued that DeepSeek stalled in early 2025 while OpenAI and Anthropic improved prompt caching, tool-calling training, and integrated deep reasoning. This is his stated opinion from the podcast, not an independent benchmark.
Is vibe coding the same as not programming?
In the way the term was used on the live, vibe coding means building applications primarily through AI assistance while understanding less of the code by hand. Igor 3K embraced the label, and the live code review showed both the speed and the messiness of that workflow.
Was the Claude Code layoff story confirmed?
No. The host presented it as an unverified rumor from chat, named no company, and repeatedly reminded viewers that his channel is gossip and speculation.
What was Akita's advice for running AI coding agents safely?
He recommended running agents inside containers or virtual machines on a separate host, such as a Proxmox server, instead of trusting an agent's own sandbox claims on your main machine.
How does prompt caching save tokens?
Prompt caching stores the computed context of previous requests so the model does not recalculate the entire conversation history each time. Akita named it the first of the three improvements separating current OpenAI and Anthropic models from DeepSeek.
What is mixture-of-experts in simple terms?
It is a model architecture that activates only the parameters relevant to the current task instead of the whole network. Akita's example: a 200-billion-parameter model does not need every parameter in memory to translate text, so it can run on hardware that could never hold the full weight set.
How much did the rumored bootcamp charge students?
Around R$40,000, according to the host's retelling, under an arrangement where students paid nothing upfront and had to prove employment with monthly documentation afterward. He named no school and presented the story as unverified gossip.
How long is the source video?
The live runs about 79 minutes. Most of the career advice and the two gossip segments come from the second half, after the Akita reaction and the GitHub review.
Turn this kind of live commentary into readable content
This episode worked because it mixed a technical explanation, a live code review, and honest labeling of what was rumor versus fact. That mix is hard to preserve in a video, and most viewers will never scrub through it the way this article did.
If you stream or record conversations like this, Skala Blog turns the YouTube video into a structured written article: paste the URL, the video is transcribed, and the transcript becomes an edited piece readers can search and cite. Your best live moments deserve a format that outlives the chat replay.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits