Impulse Tech

Scientist. Investor. Builder. @MIT PhD.
Lab → tech startup → grew a real-asset portfolio → back in tech building AI solo.

[Impulse Tech @impulsetech_au]

Everyone can vibe code now. A working app takes an afternoon. So what's left that's worth anything?

I just read a piece by Tw93, the developer of Mole (a Mac cleaner), and it's the best answer I've seen. His approach: let an AI agent install 853 Mac apps one by one, uninstall each, and check what's left behind. Then do the same for 30+ AI tools. Then hand-email 400+ users who bought but never activated.

The AI did most of the grunt work. But look at what the testing found. His first version treated the skills/ folders that skill installers drop into every agent's directory as leftover junk, and nearly deleted his own installed skills. Another case: "cache" cleanup looked safe until shared data turned out to be used by other apps. He says the places that should NOT be cleaned (30+ in the first batch) are as valuable as the new leftovers found.

That's the part vibe coding doesn't give you. A model will happily write a cleanup tool that runs, passes a demo, and deletes the wrong thing on someone else's machine. Knowing what not to touch comes from having been burned before.

So here's how I see the split:
>AI makes the code cheap, and makes tedious testing possible for one person.
>Security review and code review are still your job.
>Real-world edge cases are the other half: what happens on a machine that isn't yours, with data you didn't anticipate. (1/2)
>Experienced engineers matter more now, not less. Their job shifts from typing code to judging what's safe to ship.

When anyone can build the thing, the product is the details nobody asked for. Worth a read.

https://t.co/NVQAUgwe8Y

[Impulse Tech @impulsetech_au]

DeepSeek just shipped the whole app stack as API: web search, file parsing, URL reading, OCR, ASR, TTS, plus an open-source frontend demo. A few parameters turn it on.

The model isn't the product anymore. The app layer is, and anyone can now rebuild it for the cost of tokens. A weekend project can match what used to take a full chatbot team.

Worth reading the fine print: public-facing apps are told to enable the safety check and handle "model registration" and "AI labeling" themselves. The docs don't name the regulation, but those are obligations most global builders have never had to think about.

[DeepSeek @deepseek_ai]

🚀 The full DeepSeek App experience, now via API

Search, files, links, safety, and voice. The capabilities behind the DeepSeek App are now available through the DeepSeek API.
Enable built-in features with a few parameters, or call standalone APIs such as OCR, ASR, and TTS. Build richer conversations or recreate the full DeepSeek App in your own application.

Try our open-source frontend demo, with support for importing DeepSeek conversation history.

🔗 Demo: https://t.co/TI0Iw5qLQK
📦 Code: https://t.co/ImA0LrgUgX
📖 Details: https://t.co/t0dJxXM6Lk

1/4

[Impulse Tech @impulsetech_au]

Nous Research fine-tuned Qwen to strip CCP-aligned censorship — and published the exact recipe to reproduce it. Bold move, but it reads like trading one bias for another rather than actual debiasing, and there are credible claims a chunk of the training data was just fabricated.

The real cost isn't the culture-war argument. It's that Alibaba now has a concrete reason to slow-walk future open-weight Qwen releases — that hurts the whole open ecosystem, not just one side of this fight.

No neutral model exists. But "reproducible recipe targeting politically sensitive content, aimed at users where that carries real legal risk" is a sharper move than just shipping different defaults — and it was foreseeable how it'd land.

[Jerry Xiong @Jerry_Galaxy]

我终于知道 Qwen4-27B 和 Qwen4-Next-Flash 迟迟不推出来的原因了。

昨天阿里只放了个限时两周的 Qwen4-Max API，大家苦等的开源权重连个影子都没有。直到我看到 Hermes 团队在 Hugging Face 挂出来的这张模型卡，瞬间懂了：

他们居然直接拿阿里开源的 Qwen3.8-27B 当底座，喂了 5300 万 token，模型卡第一行就贴脸开大：在这套中国模型上强化西方价值观。

这波操作我真服气了：

硅谷大厂全闭源，美国想训个西方价值观模型，居然不得不拿中国开源模型当底座。这哪叫消除偏见，分明就是把杭州的大脑换头成了旧金山的嘴。

社区做的无审查微调早就不稀奇了，但高调植入西方价值观贴脸赚流量，这简直把用心做开源的 Qwen 团队架火上烤。

真把开源搞没了，Hermes 团队以后就对着 OpenAI 和 Claude 的闭源 API 搞价值观换头吧。

[Impulse Tech @impulsetech_au]

Independent builders who own the edge stack own the margin. Optimization isn't a lab problem—it's a distribution moat.

[Alex Cheema @alexocheema]

In June @NaderLikeLadder emailed Jensen about a crazy idea we had over dinner to make Local AI 10x better on NVIDIA hardware.

Jensen replied in 7 minutes. He saw this as important work, and approved a war room for us to work on this at NVIDIA HQ.

For a month, NVIDIA HQ was the @exolabs satellite office. We were overwhelmed by the support we got. We were back to back the first week with smart NVIDIANs coming in to our war room to help us. Creators of CUDA, Dynamo, people who built NVIDIA.

Long story short with NVIDIA we managed to improve their Local AI playbooks by 9.6x (almost the 10x goal we set out with). We’ll be publishing a whitepaper on this work.

We were just a small startup with a crazy idea. Jensen took a bet on us. Jensen cares deeply about Open Source AI and Local AI.

I don't believe Jensen is replaceable, he's a 1 of 1 human, but I hope NVIDIA will continue to aggressively support Open Source AI and Local AI, and continue to take bets on small startups with crazy ideas.

Thank you Jensen

[Impulse Tech @impulsetech_au]

Once local inference is default (not cloud-expensive), every app rearchitects. Apple's move is infrastructure, not feature.
https://t.co/8GQpgS6ODL

[Impulse Tech @impulsetech_au]

Evan's right on context window blindness, but conflating 'AI code' with 'fragile code' misses the point. The leverage moves to code review and test coverage. Bad code is bad code—the source is less important than the catch.
https://t.co/7XrXdwwKYZ

[Impulse Tech @impulsetech_au]

The actual innovation: crackle and wow/flutter emerge naturally from the physics model, not post-processing. When 'imperfections' stay after you remove artifact layers, your model is doing something right.

[Guillermo Rauch @rauchg]

Built a physical simulation of a record player with WebGL2 and Web Audio. It's *shockingly* accurate: https://t.co/MBNaTfKoai.

Instead of playing a song, it works by converting it into the physical shape of a groove. Then a simulated stylus rides those walls 384,000 times a second, and you hear the voltage that motion would induce in a cartridge.

So the crackle, the pops, the wow/flutter, the dull inner grooves are natural byproducts of the simulation. You can also "DJ" the record.

(The converted songs are in the public domain or CC-licensed. Code is linked for you to remix and re-deploy.)

[Impulse Tech @impulsetech_au]

Tests don't lie. Process (AI or manual) matters less than output. Vibe code + comprehensive tests = defensible engineering. Vibe code without tests = hoping, not shipping.

[zacker150 ➡️ TwitchCon SD @zacker150]

Actual software engineer here.

Vibe coding is when you use AI to write software without manually reviewing every single line of code. Instead, you rely on empirical tests, formal verification, and static analysis to verify that your code is correct.

Rewrites like the Discord rewrite are prime targets for vibe coding because you can easily verify them. Just take the old code, run it with a ton of random inputs (this is called fuzzing), and compare the output with the new code. Your rewrite is done when the output is always the same.

As for how you know it's vibe coded? Everything is vibe coded these days. Software engineers don't write code by hand anymore. Instead, we herd clankers writing code.

[Impulse Tech @impulsetech_au]

Claude solved a vault particle structure. Real question for credibility: what did Claude contribute—computational prediction or interpretation assistance? Methodology matters.
https://t.co/32iXe40Pfi

[Impulse Tech @impulsetech_au]

When Earth-based data centers hit physics limits, we put AI in orbit. That's not sci-fi—it's the next step when density maxes out and power hits a wall.

[Shubham Saboo @Saboo_Shubham_]

This is INSANE. A huge step for humanity.

Google Gemma 4 is a actually running at a Data Center in Space.

Read that again and let that sink in. Congrats project Suncatcher team!

[Impulse Tech @impulsetech_au]

Google misjudged parent acceptance. But backlash isn't 'no AI'—it's 'not black-box optimization of my kid's learning.' Teacher-driven AI survives this. Opaque tutoring doesn't.

[Fairplay @fairplayforkids]

📝STATEMENT: It’s clear that Google underestimated the rising parent backlash against AI and EdTech, and now it’s scrambling in the face of a PR nightmare. But this ‘clarification’ doesn’t appear to change the fact that students won’t be able to use Google Search without being subjected to Google’s AI. The company is clearly banking on Search’s popularity to lure schools into consenting to use Google’s AI.

Schools shouldn’t be fooled by this bait and switch. The good news is that it appears that Gemini, Gemini Notebook, and Search will now be off by default. Schools should leave them off, because Google’s AI tools aren’t safe for kids and undermine learning and critical thinking.

https://t.co/9XRi5LgHbH

[Impulse Tech @impulsetech_au]

OpenAI/Anthropic claim their AI broke a 27-year-old CS assumption (Exponential Time Hypothesis). Catch: the "faster" algorithm only wins past ~2⁵¹² variables — more than atoms in the observable universe. Formally verified, practically meaningless.
https://t.co/0DIlACv0aC

[Impulse Tech @impulsetech_au]

Claude running real-time fluid dynamics in Roblox. Not simulation wrappers—actual computational physics at consumer scale. Democratization isn't hype when the tech delivers.

[7vin @xounotop]

#ROBLOX #RobloxStudio #Blood #Claude #RobloxDev Real time hybrid eulerian-lagrangian fluid blood simulation running inside of Roblox

[Impulse Tech @impulsetech_au]

AI didn't reduce work. It redistributed it. Less busywork, more strategy. Whether that's a win depends on what you care about—and whether you set boundaries.

https://t.co/Fc9PQmeo4O

[Impulse Tech @impulsetech_au]

DeepSeek on DGX breaks hyperscaler lock-in. Indie builders can now compete on inference cost, not trapped in cloud pricing. The economics shift.

[Impulse Tech @impulsetech_au]

claude-seo's 26-skill suite shows the play: abstract AI into reusable toolkits, ship community-first. One person's tool engineering becomes another's platform.
https://t.co/YQsII2wovE

[Impulse Tech @impulsetech_au]

Solo founder $1M/month isn't new. AI search consolidating marketing/growth into one person—that's the real shift. Economics of leverage changed.

[Rakib Hossen @rakib_hossen_ai]

A solo founder can now build a $1M/month business without a marketing team. 🤯

AI search is changing the game.

One AI-native tool can handle growth tasks that once required an entire team.

The founders who master AI today could have a massive advantage tomorrow.

Follow @rakib_hossen_ai for more AI tools, agents & automation updates.

[Impulse Tech @impulsetech_au]

Exactly. Hand-written PHP shipped security disasters for 25 years. AI gets blamed because it's new. Root cause: misaligned incentives.

[Piyush Garg @piyushgarg_dev]

Everyone: "Vibe coding is causing all these security bugs"

Reality: PHP, Node, Linux kernel, GitLab. All hand-written code by humans, some of it 20 years old.

AI isn't creating the slop. It's finding it 🫠

[Impulse Tech @impulsetech_au]

DeepSeek v4.1 Flash 1017 > GLM 5.3 Flash > Qwen 3.8 Next Flash. Local open-source now baseline production. On-device workloads no longer need cloud APIs.

[Ahmad @TheAhmadOsman]

Top 3 Local AI models

- DeepSeek v4.1 Flash 1017

- GLM 5.3 Flash

- Qwen 3.8 Next Flash

In that order

[Impulse Tech @impulsetech_au]

DeepSeek-Huawei chip partnership isn't about AI leadership—it's about escaping supply chain hostage-taking. When everyone can train, whoever controls fabs wins. Geopolitics > architecture now.

https://t.co/VqlVFuoZKa