Dan McAteer
@daniel_mac8
I build AI agents for a living and write about it. Writing at · Writing @latentspacepod · Agentic Ops @ AnswerRocket
4.7K Following 30.7K Followers
This is utter genius.
Adaptive reasoning in Codex that updates the reasoning effort *mid-CoT* based on the task difficulty.
Powered by 'Jev' from @miu21590.
Grok 4.7 is out and its Terminal-Bench 4.0 score is horrendous.
Goes to show just how far ahead OpenAI and Anthropic are with coding models.
I still love to use Grok in Grok Bot though.
Show more
'Jev' plays Pac-Man steered by Astra.
1. Astra strategizes
2. 'Jev' carries out the strategy in milliseconds
It's early and hard to wrap my head around, but the potential is there to combine 'Jev' with larger reasoning models.
For sure.
Show more
I cannot recommend highly enough doing the following:
1. Remotely connect your ChatGPT mobile app to your laptop.
2. Go out for a walk.
3. Open the remote connection and start a voice chat.
There is something about using your actual voice to create things that involves more of you.
It's the first time I've achieved a flow state with AI.
Show more
Codexers: send THIS prompt to Codex now.
"Read Eric Provencher's post from the OpenAI Devs blog.
Audit my skills, AGENTS.md and decision boundaries and advise where I can improve them for GPT-6 Astra according to the advice in the blog."
Astra-fy Codex for yourself.
Show more
Fable 5.1 orchestrates GPT-5.6 Luna agents.
Via the 'fable-advisor' plugin for Claude Code.
1. Fable 5.1 orchestrates
2. GPT-5.6 Luna implements
3. Fable 5.1 reviews
Think of it like in-context distillation: give Luna the intelligence of Fable 5.1 at a fraction the cost.
Show more
5 Fable 5.1 optimizations you can use *NOW*.
1. Set effort to 'low'
2. Run '/claude-api cost-optimize'
3. Run 'claude-api prompt-audit'
4. Change effort mid-conversation w/o cache hit
5. Update Fable 5.1 API config w/ 'claude-api migrate'
From @RLanceMartin
Show more
Codex's upcoming 'Persistent Mode' prompt is in the Codex open-source repo.
tldr;
> A new level of reasoning effort
> Continues working until put to sleep
> Proactively generates follow-up tasks for itself
This thing is going to *RIP* with GPT-Astra, which was reportedly trained specifically for long-horizon tasks, according to The Information.
Show more
Prediction: @ssi release their first model this week.
It is a breakthrough in continual learning.
Ilya created superintelligence for real and the game is changed.
Mind blown from a new model I just got access to. I think this will be one of the most (the most?) significant drops this year. Excited. And sorry to be annoyingly vague. Just excited.
Ox-Alpha is GLM 5.3 Flash. Friends I trust confirmed it.
Its release this week will be a revelation.
In the arena on DeepSWE with the likes of GPT-5.6, Opus 5 and Fable 5, but much cheaper and open-weights?
Wow. That’s incredible.
Show more
roon says astra is a remarkable coder
can’t.
wait.
Ox-Alpha is GLM 5.3 Flash. Friends I trust confirmed it.
Its release this week will be a revelation.
In the arena on DeepSWE with the likes of GPT-5.6, Opus 5 and Fable 5, but much cheaper and open-weights?
Wow. That’s incredible.
Show more
Where humanity & AI are headed:
Opus 5 is the most misunderstood AI model.
How to get the most out of Opus 5:
1. Set effort to 'Medium'
2. Define a precise goal
3. Get out of the way
Profit.
Opus 5 is the most misunderstood AI model.
How to get the most out of Opus 5:
1. Set effort to 'Medium'
2. Define a precise goal
3. Get out of the way
Profit.
Opus 5 is the closest thing we have to whatever Anthropic has internally
it's an absolute monster if you want to optimize anything
give it a target and it will hillclimb
63% on DeepSWE is still very good, considering Fable sits at 69%. Ox Alpha will be a very popular model 🔥
Actual DeepSWE run on the ox alpha mystery model is done.
Ended at ~63% NOT the 80% my first subset test got, which makes way more sense. I've been using this thing a ton and it is definitely a very good model.
- Much better "voice" than Claude or GPT
- Decent design
- Handles subagents and long complex work quite well
- Code quality is good and it does a good job of parsing what I'm asking for
Have run into some weirdness:
- It leaves dead code around sometimes, not nearly as "through" as something like Sol
- Despite decent TPS, this thing does not feel fast at all. Especially at higher reasoning levels this thing takes forever to run, doesn't feel all that efficient. That could just be b/c I was running it in cursor, I've noticed models take longer in there and the avg tokens here seem actually pretty solid
It being on around Sol medium feels about right. This thing is definitely a good model.
The big question now is if the GLM-5.x flash rumors are true. If this really is a small model that could be run on something like 2x DGX sparks? It's gonna be a massive moment and a really big deal. Very excited to see this thing actually revealed.
Show more
AI engineers: the model absorbs the harness, and the harness transforms into an interface for human attention.
> The 'attention-interface'.
That's what 'The Evolution of the Agent Harness' is about.
A post written for @latentspacepod.
Show more
Where humanity & AI are headed:
