Vladimirs Matusevics

Senior Mobile Engineer at @bookingcom. Love to make people happy by creating Apps - @Cre8iOS, @StickeritApp

[Vladimirs Matusevics @aimorsels]

I read the AI firehose so you don't have to. 🚀

Every Thursday:
- 4 AI tools worth your time to check
- 3 good reads and what the big labs shipped
Each with a one line verdict.

Free, no spam, unsubscribe anytime.
https://t.co/FKG4Jz9vJf

[Vladimirs Matusevics @aimorsels]

Forbes asks the uncomfortable question: AI may not be a bubble, but where are the returns?

Hundreds of billions in capex. Investors want proof.

Builders see the value daily. Turning that into a business is a different problem.

Overbuilding or underdelivering?

https://t.co/5enJdGbVbj

[Vladimirs Matusevics @aimorsels]

AI costs are surging and agents use way more tokens per task. Uber blew its 2026 AI budget in 4 months. Gartner says coding costs will pass a dev salary by 2028. 💰

The fix was Chinese open-weight models. Beijing may now restrict access.

https://t.co/AGi9GZF2j9

[Vladimirs Matusevics @aimorsels]

@benjitaylor It's nice! it currently supports only Claude Code and Codex. Can more agents like Grok be added?

[Vladimirs Matusevics @aimorsels]

Grok 4.5 just dropped

Built with Cursor on real-world dev data. 80 tok/s, 4x cheaper than Opus 4.8, and free to use right now

The interesting bit is the training data. Not synthetic benchmarks. Actual coding sessions from millions of developers. That's a different kind of model

https://t.co/7cwxAFCMTX

[Vladimirs Matusevics @aimorsels]

Claude Cowork is coming to web and mobile

Unlike desktop, agent sessions run in the cloud and keep going even when all your devices are offline. Beta starts with Max subscribers in the coming weeks, doubled limits until August 5

The catch: only cloud connectors, no local tools or MCP

But Anthropic's own data explains why this makes sense. Over 90% of Cowork usage isn't coding. Roughly half is business ops and content creation. Most users never touched the local tooling anyway

They've been running doubled limits for months and just extended it another month. They really want people on this thing

https://t.co/RrcrVQ242X

[Vladimirs Matusevics @aimorsels]

New NBER paper looked at 100k+ GitHub devs plus Microsoft's internal AI tool telemetry. 🤯
Is outcome shocking or expected for you?

Code output gains are massive:
Autocomplete → +40% commits
Interactive agents → +140%
Autonomous agents → +180%

But then it all evaporates

Those 180% more commits turn into only +30% actual releases. The bottleneck? Human code review

You can generate code 10x faster but someone still has to read and approve every line

https://t.co/d602z3o9ae

[Vladimirs Matusevics @aimorsels]

GitHub trending is basically an AI agent leaderboard now. 15 of the top 25 repos this week are AI agents, LLM tools, or MCP infrastructure.

Open source used to mean libraries and frameworks. Now it means "here is my Claude Code plugin."

Is this a bubble, or is developer tooling genuinely being rebuilt around AI agents? And if the latter: what happens when the model companies ship their own versions?

[Vladimirs Matusevics @aimorsels]

4 tools blowing up on GitHub this week:

1) strix - Open-source AI pentesting. 37K stars. Finds vulnerabilities in your app before attackers do.
https://t.co/HVh08YTFk5

2) OmniRoute - Free AI gateway. One endpoint for 231+ AI providers, 50+ free. Connect any coding agent to any model. Saves 15-95% on tokens.
https://t.co/h5Ox7NUTSs

3) meetily - Privacy-first AI meeting notes. 18K stars. 100% local transcription + summarization via Whisper and Ollama. No cloud.
https://t.co/sqCscKn2Tp

4) herdr - Agent multiplexer in your terminal. 12K stars. If you juggle multiple agents like Hermes, this is a clean way to run them in parallel.
https://t.co/oKbgyCRUgK

[Vladimirs Matusevics @aimorsels]

Coding with AI doing something unexpected for engineers. It’s forcing us to become better at architecture.

When you have to explain your design to an LLM out loud the gaps become obvious. You hear them as you type. I catch bad abstractions before a single line is written now.

The funny part is I write better code even when I'm not using AI. Just the habit of clarifying my own thinking stuck.

https://t.co/yIZytnCFHA

[Vladimirs Matusevics @aimorsels]

4 open-source AI tools worth your time right now:

1) n8n - Visual workflow automation with 400+ integrations and native AI. Self-hosted, fair-code. Build AI pipelines without glue code.
https://t.co/Pxrc8zvJgV

2) Dify - Production platform for AI apps. RAG, MCP support, multi-model orchestration. Open source, deploy anywhere.
https://t.co/ARIwIqjfIf

3) Langflow - Drag-and-drop AI agent and RAG builder. Visual chain design on top of LangChain. Prototype in hours, not weeks.
https://t.co/Sis1wLMaap

4) Ollama - Run LLMs locally on your own hardware. No cloud, no API keys, your data never leaves your machine.
https://t.co/vzBXrnu3M1

[Vladimirs Matusevics @aimorsels]

@Randal_Smith Really sorry for this. I have tried to fix it in 1.9 version, hope it works for you.

[Vladimirs Matusevics @aimorsels]

I just launched ExpiresBy, a simple iOS app I built to keep track of expiration dates.
Take a photo, set a reminder, and get notified before things expire.
It works offline, and you can sync with iCloud if you want.

Would love for you to try it out:
https://t.co/1qo7iJLU2z

[Vladimirs Matusevics @aimorsels]

The AI agent arms race is getting boring.

Every launch now: "our model is the most agentic yet." Anthropic says it. OpenAI says it. Google says it. Agentic is the new "revolutionary."

But most people don't need an autonomous agent. They need reliable automation. An agent that does your job 80% right creates more work than it saves. You verify output instead of doing the thing.

The model that wins won't be the most agentic. It'll be the one that's boringly reliable.

https://t.co/LYCh8dtv25

[Vladimirs Matusevics @aimorsels]

4 CLI coding agents worth your attention right now:

1) OpenCode - 165K stars, multi-provider terminal agent. Works with Ollama, Claude, GPT. The current heavyweight.
2) Codex CLI - Apache 2.0, sandboxed execution. Best for safety-first autonomous workflows.
3) Pi - Armin Ronacher ultra-minimal harness. Lazy-loading skills, tiny system prompt. Clean architecture.
4) Goose - Linux Foundation agent. Multi-purpose automation, research, MCP integrations.

Source: https://t.co/tOrweSrgKg
Repos: https://t.co/7k4b3GXAVi | https://t.co/JXSyVwIa2T | https://t.co/HYpmtrZwtV | https://t.co/BIPU8ypDfm

[Vladimirs Matusevics @aimorsels]

My kid needed to turn a video into a Live Photo for a TikTok slideshow thing. So I vibecoded LiveMaker over a few days. Trim, pick your frame, save. All on-device, nothing leaves your phone.
Check it out https://t.co/lwSBVdxNN9

[Vladimirs Matusevics @aimorsels]

AI doesn't reduce work. It intensifies it.

Workers end up juggling multiple AI threads at once. Writing code while an agent generates an alternative. Running parallel agents. Reviving tasks they deferred for months because "the AI can handle it."

The tool doesn't give you less to do. It makes more work possible. That's a different thing.

https://t.co/SCHymWkZ0q

[Vladimirs Matusevics @aimorsels]

Sonnet 5 dropped. 63.2% SWE-bench Pro, 81.2% OSWorld, 1M context window. $2/M input, $10/M output. Better than 4.6 on coding and agents, still below Opus 4.8 and Fable on raw accuracy. This is the new workhorse for agentic coding.
https://t.co/ttouBRlsez

[Vladimirs Matusevics @aimorsels]

We are giving AI agents shell access, API keys, and file system permissions.

We are still copy-pasting emails into ChatGPT without reading them.

And we are shocked that 195 million taxpayer records walked out the door.

The agent is not the problem. The trust model is. We built tools that act on our behalf, then handed them to the same inbox we cannot stop clicking phishing links in.

That is not a security flaw. It is a design choice.

Hermes runs local-first with explicit tool gating. The difference between "agent acts" and "agent asks" is the difference between owning your AI and being owned by it.

https://t.co/oFUjrD4Ddc

[Vladimirs Matusevics @aimorsels]

4 AI tools worth your time this week:

1. Vercel Eve.
"Next.js for agents." Filesystem-first TS framework. An agent is a directory of files. Apache 2.0, any model or MCP server.
https://t.co/MqsoHHMd72

2. Databricks Omnigent.
Meta-harness above Claude Code, Codex etc. Makes them interoperable with shared policies and real-time collab.
https://t.co/El8wBa1Evo

3. Hume TADA.
MIT-licensed TTS. Zero transcription errors, 5x faster than peers, runs on a phone. 8 languages, built on Llama.
https://t.co/Xp6aVl9ORK

4. NVIDIA Nemotron 3 Ultra.
550B open-weight model pushing local AI boundaries.
https://t.co/n2bHwTCnaU

[Vladimirs Matusevics @aimorsels]

The HN thread "What was your oh shit moment with GenAI?" is full of wild stories.

One guy used Claude to reverse-engineer a 90s synth protocol in GHIDRA and had a working demo that night.

Another bricked his digital piano, so Claude decompiled the Kawai Android APK, extracted the firmware encryption key, and wrote a Bluetooth flashing script. Piano fixed in an hour.

These aren't "vibe coding" MVPs. This is real reverse engineering, protocol analysis, firmware recovery. Stuff that used to take weeks.

My question: what's the most surprising thing you've built or fixed with an AI agent?

I've been using Hermes for a while now and the gap between what people assume these tools can do and what they actually can do is still massive.

https://t.co/ID8Sd5c8F7