Executive Summary
AI coding tools and open-source models dominated tech discussions, with Fable 5 and GPT-5.6 driving attention for their capabilities in code generation and CUDA optimization. A major shift towards open-source AI is underway as models become viable for production, emphasizing efficiency and sovereignty over closed ecosystems. Meanwhile, developers are actively grappling with the philosophical and practical implications of AI, noting that while "ability is basically free," taste, prompt engineering, and tool stability remain critical bottlenecks. New hardware optimizations and local inference breakthroughs, such as GLM 5.2 running efficiently on Apple Silicon and NVIDIA's new draft heads, further signal a move toward decentralized, high-performance AI deployment.
Key Events
- GPT-5.6 Sol beats Claude Opus's CUDA speedup attainment in half the time, according to an NVIDIA Principal Engineer, while an OpenAI FDE claims GPT-5.6 Sol Ultra removes the need to read code entirely. → link
- The Hermes Agent team announces plans to expand its plugin interface, allowing developers to implement stable changes and features without waiting for core merges. → link
- Open-source AI models have reached a tipping point where they are now viable for many use cases, shifting the industry focus towards efficiency, sovereignty, and control over closed ecosystems. → link
- AI coding tools, including Fable 5, are rapidly transforming software development, though developers note that current tools still struggle to meet the stability and performance demands of a daily driver. → link
- Original LLM scaling laws were reportedly wrong due to a bug, leading to significant wasted compute, according to reshared commentary. → link
Analysis
The developer ecosystem is experiencing a clear polarization between the raw capability of top-tier AI models and the practical usability of AI tools. While models like GPT-5.6 push the envelope on performance and code comprehension, builders are increasingly vocal about the need for stability, state management in agents, and better "taste" in outputs. Open-source models are closing the gap with closed labs, prompting a strategic pivot toward local inference and data sovereignty. The coming weeks will likely see a race to refine agent interfaces—like Hermes' expanded plugin architecture—and optimize local inference stacks as developers seek reliable, customizable AI workflows.
Tweet Feed
AI Models & Research
@jezell · 2026-07-05T15:30
RT @daniel_mac8: GPT-5.6 Sol beats the CUDA speedup attainment of Claude Opus in half the time; says this NVIDIA Principal Engineer. He del… → tweet link
@nummanali · 2026-07-04T19:06
Chris is FDE at OpenAI He said with 5.6 Sol Ultra he doesn't need to read the code
Kind of gone into a trance at the thought of that Can you imagine? → tweet link
@jezell · 2026-07-04T19:54
RT @mitsuhiko: I had some vibes that Opus 4.8 was performing worse than older ones for some of uses that are off distribution and now I hav… → tweet link
@swyx · 2026-07-04T20:26
RT @sedielem: Here's a cool piece of LLM lore: the original scaling laws were wrong due to a bug, which probably led to a lot of wasted com… → tweet link
@MilksandMatcha · 2026-07-04T19:30
RT @Thom_Wolf: Most people should probably update their priors on the state of open-source speech-to-speech.
It's honestly kind of mind-bl… → tweet link
@sama · 2026-07-05T15:30
our older kid put two words together for the first time and i am approximately as amazed by this cognitive feat as i am by GPT-5.6 discovering new math → tweet link
Open Source & Agent Ecosystems
@alexocheema · 2026-07-04T21:06
It was irrational for most companies to use open-source models 6 months ago.
There was no choice then - open-source models weren't good enough.
Suddenly they are good now, for a lot of use cases.
The next year is all about efficiency, sovereignty and control. → tweet link
@TheAhmadOsman · 2026-07-04T20:53
America, 250 years of freedom is one more reason to stand behind Opensource AI and against regulatory capture narratives from corporations. → tweet link
@Teknium · 2026-07-04T21:37
Happy Fourth of July!
We want to expand the plugin interface of Hermes Agent so that many developers who have PRs waiting for very long periods can implement stable changes that they can share and publish without having to worry about getting their feature etc merged and want your suggestions on where to expand the interfaces.
With plugins, you can implement features, fixes, security layers, whatever you can think, so long as the interface for those plugins exists.
If you have ideas for an expanded plugin interface that should be implemented, please reply in the thread of this post!
We will see what we can do on that front over the next week! → tweet link
@victormustar · 2026-07-05T17:21
RT @Meituan_LongCat: 🐱 LongCat-2.0 is now fully open-source — MIT licensed, no restrictions.
Since our launch a few days ago, the response… → tweet link
@Ex0byt · 2026-07-05T04:20
Single Spark, DSV4-Flash (full weights), sglang- You said push.. I pushed. https://t.co/D0vz4ngSBh → tweet link
@KingBootoshi · 2026-07-05T06:32
i am bearish on subagents
i am bullish on specialized agents that maintain state with their own memory working with each other
ephemeral agents are a waste of tokens (besides maybe, researching or scoping)
ultra code for example is a complete waste of inference imo! → tweet link
@alexocheema · 2026-07-05T18:47
we're going to scorch the earth with high quality, open agent traces.
no moats for the closed labs. → tweet link
AI Coding & Developer Workflows
@jxnlco · 2026-07-04T22:19
transcribed notes on taste:
If you want taste, you're gonna have to eat.
I think now, as AI has allowed more and more people to create things, the biggest differentiation, the biggest skill people are lacking, is taste itself.
What does that mean, right? Taste is a little bit of aesthetic judgment, knowing what's beautiful. But I think taste is really your ability to model out what people will like. And a lot of that requires you to not just consume, right? Like if you want to be a good chef, you have to eat at people's restaurants. You can't just look at the menu.
It's curation versus consumption. And with AI there's so much more volume of work now, so it's really easy to regress to the mean. In which case it's not really taste. Taste is a kind of risk-taking. You're choosing to deviate from the safe average. → tweet link
@MengTo · 2026-07-05T09:35
We might be entering an era where writing an incredibly detailed prompt matters more than building the landing page itself.
With Fable 5 becoming this capable and this expensive, the prompt is quickly becoming the most valuable asset.
This interaction was one-shotted. The level of polish and interactivity is insane. → tweet link
@thdxr · 2026-07-04T19:40
a lot of ai coding tools ours included have not been clearing the bar for stability and performance you should demand of a daily driver
james is focused on fixing that and there's some novel things we can do to clear that bar more than we ever have → tweet link
@jxnlco · 2026-07-05T15:33
He replaced most of descript with OpenAI transcription api and ffmpeg and codex → tweet link
@jxnlco · 2026-07-04T20:26
easiest way to this is to send codex a screenshot and tel lit to use image gen → tweet link
@jezell · 2026-07-05T02:07
RT @skeptrune: the "software engineering" i was doing in 2022 has been fully killed by fable
for a minute i thought it wasn't that differe… → tweet link
@jezell · 2026-07-04T21:35
Pretty neat:
https://t.co/c8V94PsI0l
But under the covers it's actually duroxide:
https://t.co/3lr1WadVGI
Going to have to look more at duroxide in a bit. A nice embeddable rust workflow engine could certainly be useful. → tweet link
@nummanali · 2026-07-05T09:12
People have said Fable 5 is too eager but it’s all about framing
h/t to Elliot for referencing the prompting guidelines - that’s the secret sauce
TLDR - reduction of 80% on most prompts
I equate it to principle level instructions
https://t.co/kCGKjaeoU2 → tweet link
@ivanfioravanti · 2026-07-05T17:02
GLM 5.2 running in Hermes Agent commenting code written by GPT 5.5... "The extraction function is a mess" 🤣 https://t.co/iR8M6YBpKk → tweet link
@juliarturc · 2026-07-05T17:52
Imagine if Apple said: Your iPhone takes crappy pictures because you're using it wrong. Most people just press the camera button.
You need to do 3 jumping jacks, hold your iPhone at 12.5 degrees and whisper a mayan incantation to Siri so that she can call the tool that takes a picture.
Also, please stay tuned for updates to these shenanigans every 3 months or so. → tweet link
Hardware & Local Inference
@ivanfioravanti · 2026-07-04T20:20
Here it is! GLM 5.2 4bit running on a single M3 Ultra 512GB at ~16 toks/s in a video, while ds4-eval video is on q2 version ~17 toks/s. Test in progress! https://t.co/QZYlbiyWEZ → tweet link
@ivanfioravanti · 2026-07-05T12:43
I discovered just now that @NVIDIAAI released 2 Dflash draft heads 3 days ago on HF! I missed this move! Thanks 🙏
https://t.co/8QuF77XZZl https://t.co/J3nRNVg6YF → tweet link
@gospaceport · 2026-07-05T06:43
Time to pull out some Optane and pSLC Phison specials I had them cook up for me a few years back 👀 HT @LebanonJon for that hookup 🔥 https://t.co/vyZRgO3vYf → tweet link
@ivanfioravanti · 2026-07-05T13:42
Many people asked me what the status of mlx-lm is. A picture is worth a thousand words.
Main MLX library is a different story.
I hope Apple will invest more resources in this team, because they built real magic so far. https://t.co/wiCTaeJAHw → tweet link
@ivanfioravanti · 2026-07-05T18:33
Another experiment on ds4 engine 🚀 GLM 5.2 batch inference for ds4-eval only. It's still a draft, but it seems working.
Sequential ~16 toks/s Batch 8 ~19 toks/s (in the video is 23 but later on it stabilizes to lower rate it) Batch 12 ~21 toks/s Batch 16 ~ 22 toks/s
I'll let them b12 finish to compare results 💪 → tweet link
Industry Trends & Developer Commentary
@FinansowyUmysl · 2026-07-05T08:44
Świetna diagnoza.
Dodałbym, że poza poniższymi czynnikami są też czynniki zewnętrzne:
Koniec ery taniego pieniądza (wysokie stopy) wymusza na firmach technologicznych zwiększania marży. Kiedy stopy były prawie ujemne, inwestorzy byli skłonni inwestować w startupy dużo, bo nie mieli alternatywy... AI to kolejny czynnik, ponieważ spowolnienie w IT zaczęło się jeszcze przed masowym wdrożeniem AI. Samo AI spadło korporacją z nieba. → tweet link
@steipete · 2026-07-04T19:03
In the next version of https://t.co/B1RkFJhhSH - see exactly when you resets expire so you can up your ̶t̶o̶k̶e̶n̶m̶a̶x̶x̶i̶n̶g̶ valuemaxxing game! https://t.co/04q26JVtiC → tweet link
@ivanfioravanti · 2026-07-05T17:20
For anyone out there using Microsoft dotnet and coding agents, this is your repo: https://t.co/i0KNPDRjS3 → tweet link
@badlogicgames · 2026-07-05T18:12
fable sucks at finance. hard. → tweet link