Teddy

⚔️ AI model comparisons & one-shot tests
🛠️ Exploring tools, building with AI
🧠 AI news, experiments & discoveries
🏭 Grok bot enthusiast

[Teddy @teddyaisi]

A convenience store at 2 AM. Which one feels the loneliest?

Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5.
Three takes on rain, light, and reflections.
Which would you pick on visuals alone?
Now check the time and cost. Same answer?

[Teddy @teddyaisi]

Nice little fix in Zed 1.26.2: ACP defaults now only apply to new conversations. Reopen an existing chat and its settings stay put.

Picking up an unfinished task shouldn't come with another round of checking your settings.
https://t.co/kBMWiy4zTE

[Teddy @teddyaisi]

The Information reports that DeepSeek and Moonshot trained models using Astra outputs. That makes training-data provenance worth watching, but it doesn't tell us which capabilities came from those outputs.

https://t.co/HIPmsrdQpC

[Teddy @teddyaisi]

A useful detail in MCP reference servers v2.2.0: the Everything server blocks non-public addresses throughout fetch redirects.

Checking the starting URL alone leaves later hops unchecked. Glad to see the protection follows the redirects too.
https://t.co/x44EZZSfiE

[Teddy @teddyaisi]

llama.cpp v0.8.0 adds LFM3 support, plus realtime audio and text-to-speech endpoints. The voice endpoints are the interesting bit—less plumbing to build yourself for a local voice app. The streaming-audio API is still experimental, though.

https://t.co/Txm3mXAePe

[Teddy @teddyaisi]

Qwen Code v0.26.3 lets generated UI cards keep their inputs and state when you navigate away and back.

Small thing, but having to fill everything in again would get old fast.

https://t.co/zl1ItS1DFI

[Teddy @teddyaisi]

Less time fixing peaks sounds like a useful job for AI. Waters says its new Empower peak integration uses models trained on a lab’s own data, with results analysts can review.

Early access starts in December 2026; validated deployment is planned for mid-2027. The useful part would be fewer routine corrections without making results harder to check.
https://t.co/ZXLEIuVRlb

[Teddy @teddyaisi]

NVIDIA’s new Nemotron 3.5 Transcribe is open-weight and transcribes 40 languages.

But the detail that caught my eye is domain vocabulary tuning.

Getting specialized terms right sounds more useful than adding another task to the feature list.

[NVIDIA AI @NVIDIAAI]

Meet NVIDIA Nemotron 3.5 Transcribe 🎙️

An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.

Watch it keep track of who said what when everyone’s talking at once.

1/ 🧵

[Teddy @teddyaisi]

@alexandr_wang Do not bother Tibo and poteto.

[Teddy @teddyaisi]

@QwenAIFans Really!? What level do you expect it to reach?

[Teddy @teddyaisi]

A glowing deep-sea jellyfish, brought to life in 3D.

Is Terra good enough?

[Teddy @teddyaisi]

@Its_lakshya_ai @OpenAI True. The issue continues to arise across months.

[Teddy @teddyaisi]

@OpenAI when the FUCK are you fixing this?

[Teddy @teddyaisi]

@kimmonismus Yeah it’s insane. Hope it will be available to everyone soon.

[Teddy @teddyaisi]

@elonmusk Both highest and lowest lol

[Teddy @teddyaisi]

@thsottiaux I still don’t get the reason why ChatGPT work exists. Codex does everything you want.

[Teddy @teddyaisi]

@sundarpichai @ArtificialAnlys Please be available for everyone soon.

[Teddy @teddyaisi]

@elonmusk insane Google is back

[Teddy @teddyaisi]

One geode, three interpretations.

GPT-6-Astra, GPT-6.1-Sol, and GPT-6.1 Terra,
same prompt, high reasoning effort, one generation each.

Focus on amethyst facets, gallery lighting, and a slowly turning display.
Which one do you like?👇