spacy

LLM research, systems and compilers | ax + dspy in TS | agent engineering.

[spacy @dosco]

we're on track to building the best dspy powered llm framework in the typescript world

[Omar Khattab @lateinteraction]

@VarunGodbole @isaacbmiller1 @DSPyOSS Yes — @dosco’s Ax is pretty much the official DSPy in TS.

[spacy @dosco]

sneak peak into the next dwarkesh pod.

[spacy @dosco]

love codex, random rewards is their best feature

[Tibo @thsottiaux]

Hi. Over the last 24 hours we had three separate small incidents that affected Codex reliability. Those are three too many and we are taking active steps for them to not reproduce.

I have reset usage limits for Codex across all paid plans. May the tokens flow again.

[spacy @dosco]

@robjama is that broadway hotel, lesliville. toronto is a really nice place in the summer but vans still hard to beat waiting on toki to fix the tech hangout scene out here.

[spacy @dosco]

google does so much good. their awesome small models also probably help them with data work for their big models.

[Google @Google]

Today we’re introducing Gemma 4 12B — our latest open model that brings advanced agentic reasoning, vision and audio directly to your laptop.

It delivers performance nearing our larger Gemma models with a much smaller total memory footprint, while being small enough to run locally with just 16GB of VRAM. It’s open and accessible for everyone to use under a permissive Apache 2.0 license.

This is all made possible by our new, unified architecture that removes separate multimodal encoders. Here’s how we did it 🧵

[spacy @dosco]

ax has audio support for exactly this reason going ahead agents and audio is gonna dominate.

[Aoden Teo @AodenTeoMT]

Today, we’re excited to introduce Miso One, the most emotive voice model in the world.

Miso One is an 8-billion-parameter text-to-speech model for highly expressive speech generation. It emotes like a human and responds faster than a human, with just 110 milliseconds of latency.

We’ve open-sourced the model weights, with API access coming soon.

Hear how Miso One sounds in the thread below.

[spacy @dosco]

@NoahZiems i have a couple model ideas i'm exploring one sorta a DSPY+RLM tune with knowledge of the https://t.co/dDHZsiZw7z harness. hook me up :)

[spacy @dosco]

@tokifyi 🙏

[spacy @dosco]

testing the https://t.co/dDHZsiZw7z agent for all these models it works with 4b qwen so these will kick ass and save everyone some token costs

[Bindu Reddy @bindureddy]

Open Source AI has gone up by 500% in 2 months

- Gemma 4 12B just dropped
- Deepseek Flash is used in production workloads
- Kimi 2.6 for coding
- Minimax in always-on agents

Bigger growth rate than Anthropic 😉 🚀

[spacy @dosco]

https://t.co/nAVc2uNweQ

[spacy @dosco]

todays reading

[spacy @dosco]

@BrandonLeafman no it just gives you a working agent. it comes with .optimize() method that uses GEPA you have to setup eval etc

[spacy @dosco]

use perplexity, parallel, google, x search whatever and build this in 5 minutes using DSPy+RLM (ax-agent) https://t.co/y45nH1PCrQ

[Perplexity @perplexity_ai]

Introducing Search as Code, our new search architecture for AI agents.

It writes Python that calls our search stack directly, instead of looping through function calls one at a time.

Available in the Perplexity Agent API, and now default in Computer.

https://t.co/ut6GGWQTVO

[spacy @dosco]

this is very cool, especially the idea that the final artifact is for a machine, maybe a tree of these agents could work on a larger problem passing these artifacts as zip files between them,

[Han Xiao @hxiao]

Sharing a project I've been heavily using - Dataroom. It's a local-first harness that runs deep research with a small language model and gives a zip file at the end. Deep research is becoming an important first step for long-horizon tasks (the 2nd step being implementation), and I believe a small local model in a disciplined harness handles it well - we shouldn't waste frontier-model tokens on it. Dataroom runs on your own GPU at near-zero marginal cost, and it can keep going for hours until the dataroom is genuinely comprehensive, instead of stopping when a metered budget runs out.

[spacy @dosco]

be careful out there the phishing attacks are getting smarter trust no one

[spacy @dosco]

if you studying cs then study CS+LLMs, important to be good at your thing and also LLMs. systems thinking comes from knowing more stuff

dist. sys. - consensus algos, lsm-trees, trade offs
os - cgroups, namespaces, ebpf, firecracker, gvisor
compilers - llvm, mlir
languages - c++, cuda

[spacy @dosco]

💯 i find my deep knowledge of distributed data systems, compilers, various languages, managing large code bases massively compliment my ability to vibe with the models, i'm able to jump ahead of issues, provide guidance when the plan is too naive, help structure the whole system. one thing thats not changed from the pre-model days is i'm still holding several balls in the air in my head just that now they're high level balls not low level bs.

[kache @yacineMTB]

We are in a golden age where, if you are good at systems and understanding, AI increases your abilities by an order of magnitude. But if you are not good at it, you just spin your wheels and end up nowhere helpful

[spacy @dosco]

humans teachers are people that we lookup to and are inspired by until we grow up and find new mentors when we look to specialize this seems to be the golden spot for LLMs, however good the professor he just does not have the patience and time most people need to gain a deep intuitive grasp of complex topics the LLM has no problem it’ll answer every one of your million questions and more

[Dwarkesh Patel @dwarkesh_sp]

Over the last 200 years, we've automated away a lot of hard physical labour. But people still go to the gym.

Indeed, many people today are more physically capable than people in the past. We can train systematically for whatever physical goal we want, and it's more fun than hard labour on a pre-modern farm.

@karpathy's hope is that, in the future, the same will be true of learning.

AI tutoring that's tailored to each person will make learning easy, and more people will want to do it. We will be able to go much further than our ancestors.

[spacy @dosco]

@yeguacelestial @yacineMTB models often take lazy route, rush, over complicate, can’t see the big big picture… the list goes on you need a biggest model out there to provide guidance for now that is human experts

[spacy @dosco]

@geteviapp i fell for the RLM useful to me velty...

[spacy @dosco]

starting a challenge put your explanations below.

[pracurser @PracurserTweets]

@dosco Sowy.. but can someone explain RLMs to me pls ty 😭

[spacy @dosco]

@Catchingtides yup DSPy is all about the composability