NZ retweeted
Before telling my wife I need to buy 8 more GPUs
I practiced on an AI version of her first.
She pushed back hard.
I got cut off.
I said "yeah" and she kept going.
I paused mid sentence and she waited.
Felt way too real.
Built with Fish Audio's Live 3
It listens and talks at the same time.
The practice call went somewhat okey.
Now time to do it live.
Wish me luck.
🤝 Paid partnership
Introducing @FishAudio Live 3: a full-duplex voice model that listens, says "mm-hmm," and knows when to shut up.
Enterprise? Book a demo for your free custom agent: s.fish.audio/vxhgs6
Reply with your use case - we'll build it free for the first 1,000 people.
NZ retweeted
paying for a server just to wait on a reply doesn't sense
your agent now has a bedtime
Inkbox gives agents their own email, phone, iMessage and WhatsApp,
and starting today it hosts them too:
> it sleeps when idle, so you don't pay for downtime
> it wakes up the second someone replies
sleeping when idle is underrated
🤝 Paid partnership
We raised $5.3M to give AI agents a computer to run on, with their own email, phone, iMessage, and WhatsApp
Led by @peakxvpartners, with @rauchg, @sternhenri, @jefrankle, @cjc, and @daniel_levine
Today we’re launching hosted agents on @inkbox_ai. Your agent runs next to its communication channels, so it never misses a reply. It sleeps when idle, so you don’t pay for downtime.
In the 4 months since our YC launch, we’ve processed millions of messages for AI agents.
NZ retweeted
Frontier performance at a fraction of the compute.
Today at @i14labs we’re launching Sugar Glider, our new family of LLM architectures!
Join the waitlist for early access: i14.ai
Sugar Glider delivers a new Pareto frontier: extreme memory and compute efficiency whilst maintaining frontier performance, even for large models (tested up to ~1T params).
O(1) per token inference → ✅
Fixed size KV cache → ✅
Frontier long-context retrieval quality → ✅
1 billion token long context retrieval → ☑️**
(**100% -> single needle @ 1B-token context, 90.5% multikey 1 @ 1B-token context.)
How did we do it? Turns out a few decades in the trenches of human memory and language model research is a pretty handy starting point to solve the “intelligence vs efficiency vs memory” conundrum.
... and just for something extra, we’ve been converting some of the most powerful and popular open-weight models to O(1) inference and fixed memory, with each conversion taking just hours :)
Tune in via our waitlist for early model access i14.ai
More results dropping soon!
NZ retweeted
Claude just dropped fable 5.2 and i asked it to build a japanese floating island
same price as 5.1. better model.
in the video:
- the benchmarks
- the official launch
- the floating island demo i built with it
watch till the end
NZ retweeted
We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.
NZ retweeted
Claude Mythos 5.2 is out.
Somewhere a junior dev just updated his LinkedIn to "AI orchestrator"
NZ retweeted
Your Mac's clipboard remembers one thing. So I built Pocket 👇
Hover the notch → everything you copied drops down: text, links, screenshots, files. Search even reads the text inside screenshots.
Free + open source:
github.com/minutechreview/po…
NZ retweeted
Bambu X2D just printing music by Claude xitter.twitterapi.xyz/i/broadcasts/2rJqarpMV…
NZ retweeted
Your AI assistant has access to Gmail, Google Drive, Slack, Notion, and a dozen other tools.
Yet it still struggles to complete simple tasks.
Why?
More integrations don't automatically mean better results.
Here's how to make your assistant more useful
NZ retweeted
claude code got a lot better in october
anthropic shipped mods, and the desktop app finally handles a bunch of sessions at once. i can have one fixing bugs and another writing tests while i just check what they did
you can also see what's eating your usage now, which i really needed
they keep shipping and the tool keeps getting better every month
NZ retweeted
OpenAI keeps copying Anthropic , and it's getting hard to ignore 👀
> Codex and ChatGPT now share one usage limit (rolling out fully in December).
Claude Code and Claude Chat have always worked that way.
> Canvas launched a few months after Claude's Artifacts.
> Codex CLI dropped a couple of months after Claude Code.
> OpenAI added full MCP support, a standard Anthropic created and open-sourced.
Not hating. Competition is good for us. But the pattern is getting hard to miss.
Are we watching OpenAI copy Anthropic's playbook, or is this just where the whole industry is heading?
NZ retweeted
Does OpenAI not know that the usage limits are one of the main reasons people still use ChatGPT?
A lot of people fw GPT because of the long limits and image generation.
I'm sorry OpenAI, but this is going to go sideways for you guys.
NZ retweeted
We've raised the max request size for Messages and Token Counting on the Claude API from 32 MB to 128 MB.
Image-heavy requests like long computer-use sessions and big PDFs were previously hitting 32 MB with plenty of context window left. Now you get ~4x more room.
NZ retweeted
First impressions of Gemini 4 Carbon’s preview by @GoogleDeepmind are in. @petergostev just put it through a long run of 3D generations, creative prompts, visual explainers and UI tasks, compared against Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol, Muse Wonderful, Kimi and DeepSeek.
Watch now to see an early preview of the latest model coming from @GoogleDeepMind.
0:00 Intro: Gemini 4 Carbon
0:15 First test (White House): Carbon vs Argon
1:28 Carbon vs Claude Opus 5.5
2:42 Carbon vs GPT-6.1 Sol
3:06 Carbon vs Muse Wonderful and Kimi
3:48 Why test with 3D generations
4:45 Playable 3D game: Carbon, Argon, Sol, Opus
6:28 Where Carbon falls short: gaps in detail
7:50 Stonehenge
8:30 Golden Gate Bridge
10:41 Golden Gate Bridge: GPT-6.1 Sol comparison
11:53 Carbon's creativity
15:33 Monet paintings transition
18:37 Van Gogh paintings world
21:47 Why creativity matters in a model
23:30 Another 3D prompt: Carbon vs Opus, Sol, Muse, DeepSeek
26:27 London through the ages
29:56 Istanbul
31:44 Harbin ice city
33:20 New test: visual explainers (how lightning works)
35:34 John Snow and the cholera map
38:20 UI work
39:45 Golden Gate Bridge completion animation
40:43 More UI generations
41:55 Ski resort site and ramen ordering
44:20 Verdict
NZ retweeted
🚨 A "gemini-3.8-flash-router" slug has reportedly been sitting in Antigravity's metadata for a couple of days.
I'm increasingly convinced the next Flash ships as a router, sending easy prompts down a cheap path and hard ones to a heavier model.
Maybe it's just an internal test name, but right now that's the cheapest way for Google to look fast and smart at the same time.
Gemini 4 Flash might be in grey testing for some users..
And its pretty fast while being extremely close in capabilities with Gemini 4 Argon.. could be even better when its released..
Carbon is looking pretty good too from what we're seeing on arena..
I'm not getting any more outputs of it since yesterday.. so a new checkpoint might be coming soon..
NZ retweeted
Today, I'm launching llmdojo.dev:
- The hands-on way to learn AI engineering: LLMs, RAG, agents, evals and inference.
- 50+ interactive labs. Tweak tokenizers, attention, embeddings and KV caches live.
- Train a neural network in your browser and watch it learn.
- Build agentic workflows with tool calls, memory, guardrails and bounded autonomy.
- Ship RAG that works: chunking, hybrid search, reranking, and LLM-as-a-judge evals.
- Production AI: model routing, fallbacks, quantization, caching and cost per request.
- 490+ interview questions, reported from OpenAI, Anthropic, Google, Meta, Amazon, Microsoft, Mistral and 50+ more.
- Every answer is cited from papers and books, or measured on real models.
