AI engineering, cybersecurity, local LLMs, ai tools, open source
Joined June 2013
- Tweets1.4K
- Following624
- Followers758
- Likes3K
Pinned Tweet
Local Qwen generates 234 tok/s in JSON format. In plain text, only 70
The same local AI model: 70 tok/s in prose and 234 tok/s when generating JSON.
More than 3 times faster. Without changing hardware.
In Qwen3.6-35B-A3B-Splash tests on Mac mini M6:
- EN Prose: 69.6 tok/s
- Python Code: 135.9 tok/s
- JSON Extraction: 234.1 tok/s
Why? Speculative decoding works better when subsequent tokens are easier to predict.
Repetitive JSON and code structures can provide much greater speedup than ordinary conversation.
xitter.gq/swarogan/status/210392…
Before AI transforms your company, it might destroy what already works.
When we introduce AI into a company where data is poorly organized, it begins to act like a virus, replicating and amplifying everything that is wrong and introducing even greater chaos.
xitter.gq/swarogan/status/209633…
Swarogan retweeted
Gemini 4 Hydrogen vs. Gemini 3.8 Flash: Motion Graphics Comparison
The speed gap is wild
For context on the setup, Gemini 4 hydrogen in battle mode went head-to-head against 3.8 flash and also ran Gemini 3.8 flash separately in direct mode:
generation times:
Gemini 4 hydrogen took about 2 minutes
Gemini 3.8 flash took about 11 minutes
Swarogan retweeted
Hi guys, genuinely curious about what I’m seeing in the field right now:
1. Why are people canceling their Apple M5 512GB orders and showing they got the RTX Pro 6000 GPU instead?
2. And if some people are canceling, how can my line move faster so I can get the Apple M5 512GB earlier than 2027? 😅
Curious what everyone thinks - pure performance, software ecosystem, availability, or something else?
#MacUltra #RTXpro6000
35k ⭐ on GitHub in a month for skills.
Here are the 10 fastest-growing Claude Skills repositories in October 2026.
Among the leaders:
Matt Pocock Skills: +35,301 ⭐
Ponytail: +32,647 ⭐
Security Audit Skill: +31,357 ⭐
Security audit, code analysis, interface design, technical diagrams.
Which Skills do you use?
If you're building games with AI, you need this!
I'm experimenting with creating my own MMORPG using AI, and one of the biggest problems is constantly checking whether the generated code actually works.
I just found Genex Desktop, an open-source project that solves this.
The agent doesn't stop at writing code. It can run the game, control a character, observe gameplay, detect bugs, and go back to fixing them.
All of this works with Godot, Unity, or Unreal Engine, with the option to use local AI models as well.
If you're making a game, show me what you're working on!
Swarogan retweeted
October ❌ Hacktober ✅
- PHP got hacked
- Railway got hacked
- OpenCode got hacked
- Binance got hacked
- Polymarket got hacked
- GitHub Actions got hacked
- SoftBank got hacked
- Cursor Marketplace got hacked
- Coinbase vendor got hacked
- Fireworks AI got hacked
- LiteLLM got hacked
- Capcom got hacked
- npm packages got hacked
- TiDB Cloud got hacked
- Warp's vendor got hacked
Let's see what November has in store for us 💀
Swarogan retweeted
Am I the only one that is excited about Thinking Machine?
Models meant to be fine-tuned and specialized seem like the play, given the major players are battling with losses.
Probably even better with the rise of LLMs rewriting weights. Feels like it will be strong in a few years, esp. as hardware catches up.
The previous Inkling Small model's size is only 276B total, 12B active.
Hope they release additional Mini or Nano models soon or later.
You can try Inkling 2 Small on Tinker Playground today. The model is also available for fine-tuning and sampling on Tinker; checkpoints will be downloadable with the release of Inkling 2 Small’s weights.
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/ink…
Swarogan retweeted
I jinxed so hard that, they have to release one day early with more than one model.
Seems I have Jinxed super power.
Anthropic guys if you need it, just let me know. I can help Fae 5.5 drop early.
This niche has just closed for you!
If you're a programmer who thought you could escape AI's takeover of coding by moving into cybersecurity, AI is already there.
OpenAI has just released GPT-6 Cyber, which discovered 81 previously unknown vulnerabilities in Firefox and developed working exploits for 948 vulnerabilities in the Linux kernel.
Swarogan retweeted
Everyone keeps asking the same question: Why hasn’t Microsoft built the model that beats OpenAI, Anthropic, or Google?
I think that question misses Microsoft’s real game.
Imagine the AI race as a gold rush. Some companies are fighting to discover the biggest gold mine.
Microsoft is building the roads, selling the shovels, renting the machinery, owning the land around the mines, and now quietly digging a few mines of its own.
That’s the strategy.
Microsoft already owns the distribution layer:
- Windows
- Office
- Teams
- GitHub
- Azure
- Microsoft 365
Hundreds of millions of people and businesses already live inside that ecosystem. So Microsoft doesn’t necessarily need the world’s #1 model.
It needs the world’s best models, whoever builds them, running through infrastructure and products Microsoft already controls.
And that’s exactly what it’s doing.
- OpenAI models? Available.
- Other frontier models? Available.
- Microsoft’s own MAI models? Now coming too.
So the game is becoming:
model intelligence → Azure / Foundry → enterprise data → Copilot → users
The model becomes one layer. Microsoft owns almost everything around it.
That is why the numbers matter.
- Microsoft Cloud crossed $214B+ annual revenue.
- Azure grew 41%.
- Microsoft 365 Copilot passed 30M paid seats.
The company is already monetizing AI before it has to win the frontier-model leaderboard.
And this is the part I find most interesting:
- If OpenAI wins, Microsoft can benefit.
- If another model becomes better, Microsoft can offer that too.
- If Microsoft’s own models become competitive, margins and control improve even further.
That is a very different position from a company whose entire advantage depends on one model staying ahead.
Maybe Microsoft isn’t late to the AI race at all. Maybe it chose a different race.
The real question is:
Does Microsoft need to build the smartest AI model in the world… if it can own the place where the smartest models get used?
That might be the more important battle.
Follow @SiftByte for more on how the companies, products, and infrastructure shaping the next era actually work.