achilles

[achilles @killthebuddha_]

agents are the wrong abstraction

[François Chollet @fchollet]

Sufficiently advanced agentic coding is essentially machine learning: the engineer sets up the optimization goal as well as some constraints on the search space (the spec and its tests), then an optimization process (coding agents) iterates until the goal is reached.

The result is a blackbox model (the generated codebase): an artifact that performs the task, that you deploy without ever inspecting its internal logic, just as we ignore individual weights in a neural network.

This implies that all classic issues encountered in ML will soon become problems for agentic coding: overfitting to the spec, Clever Hans shortcuts that don't generalize outside the tests, data leakage, concept drift, etc.

I would also ask: what will be the Keras of agentic coding? What will be the optimal set of high-level abstractions that allow humans to steer codebase 'training' with minimal cognitive overhead?

[antirez @antirez]

Codex seems semantically more powerful than Claude Code. Maybe Claude is a better overall agent, a lot more "respectful" of the files, less prone to do a mess. But GPT5.2 is more capable. I have no doubts... I wonder why Claude Code is *so much* more popular.

[Daniel Jeffries @Dan_Jeffries1]

Many people cannot think clearly, specify what they want exactly, or explain it properly.

That means AI is of zero use to them.

[eric zakariasson @ericzakariasson]

opus 4.5 is the model for sync work
gpt 5.2 is the one for async

[achilles @killthebuddha_]

software manufacturing

[achilles @killthebuddha_]

software manufacturing

[The Broken Binding | Bookstore @binding_broken]

We are beyond thrilled officially reveal our special edition of The Devils by @LordGrimdark himself! 😈

[achilles @killthebuddha_]

The research preview of #ChatGPT #operator mode is pretty mind blowing.

[Yasser @yasser_elsaid_]

I miss text-davinci-003 so much

[ARC Prize @arcprize]

ARC Prize is now 3 months old - we're announcing:

🏆 +$100K Grand Prize (now $600k)
📜 +$25K Paper Awards (now $75k)

And we're committing funds for a US university tour in October and the development of the next iteration of ARC-AGI.

https://t.co/35KOPOJG9L

[achilles @killthebuddha_]

@arcprize there's something important missing here. o1-preview is a step-change improvement when given just a little bit of feedback. for example, prompting like https://t.co/rwL9mQSkEj fails for every other model. imo this is more than just "more compute, more accuracy"

[ARC Prize @arcprize]

We put OpenAI o1 to the test against ARC Prize.

Results: both o1 models beat GPT-4o. And o1-preview is on par with Claude 3.5 Sonnet.

Can chain-of-thought scale to AGI? What explains o1's modest scores on ARC-AGI?

Our notes:
https://t.co/sV6LM1foGx

[Shane Mac @ShaneMac]

Someone should deploy this with an ENS name on XMTP so we can build apps inside messaging w/ native financial rail and then use it instantly with friends in group chats.

Never leave @converseapp_ - build apps, use apps w/ friends, repeat.

[achilles @killthebuddha_]

This can't be the simplest way to do a "safe" intersection between object types in @typescript...right? 90% sure I'm missing something here.

https://t.co/kLnMAyfFuu

[XMTP @xmtp_]

Join the next Dev Call to learn how to build & test bots w/ XMTP + MessageKit. 🛠️🤖

Building mini-apps w/ XMTP for #ETHOnline? This call's for you! 🫵

Bring your❓& let's learn together! 🫡

When: Tue Sep 10 @ 9 am P
Where: https://t.co/CvhidLYP2g
RSVP: https://t.co/UPGagKifVs

[🍓🍓🍓 @iruletheworldmo]

rushed a little but will refine and add some more info I've been given if it bangs.

-project strawberry / qstar

ai explained has been close to this for a while so i'd watch them for a cleaner take if you want to dig in. this is what ilya saw. it's what has broken math benchmarks. it's more akin to rlhf than throwing compute at the problem. sus column r is a very very tiny open ai model using strawberry. strawberry in the larger models comes on thursday.

think of it as an llm fine-tuned to reason like a human. hence why sam liked the level two comment, and felt great about it. ilya did not. here we are.

-huge models, sora, voice, video and safety.

i'd referenced some model sizing based on meta and claude having small 8b, medium 72b and large 405b. this is a simple way to frame and means nothing. except that a much larger version of 4o is coming. when you try it, it will be the first noticeable jump that we saw when going from gpt 3 to 4. the jump from original 4 to sonnet 3.5 will seem insignificant in comparison. arrives next week with strawberry.

gpt next. etc.

so gpt next (internally called gpt x, you could call it gpt5) is also ready to go. lots here relies on safety and what google do next. it's difficult to say if competition will trump safety.

though red teaming is finished and post training is done. this model is such an enormous leap in capabilities it's becoming impossible to make the model safe. if you had this particular model unlocked, you could easily disrupt the world on an unprecedented scale. when you mix in voice, video, sora, agents, and the eye-watering capabilities, things hot up. they'll get the safety right and they'll roll it out I'm sure.

this is why we post don't die or vague post around how everything is about to change forever etc. it is. we've tried the models. it's insane. i'm not directly an agent, though i've had access to an early benchmark of five to take over an account and influence some big names in the field to carry out a few things for me. github was one such case of using the model to convince several to launch.

sora and voice rollout
it's expensive. especially sora. it's proving incredibly difficult to make safe. without guardrails for example you can with a simple prompt create a video of a world leader saying anything in their own style and voice, and effortlessly hack into large scale state secrets. if you haven't read situational awareness, it lays a lot of this out.

we will get a step change next week

it won't quite be gpt5. gpt5 / next / x / is more comparable to the jump made from gpt1-4. this is why sam feels great. ilya was right. you can scale your way to a digital god with or without strawberries. but strawberries + scale will cure world problems overnight.

sam. obviously not random chance you'll see i've been rocking with current / former openai employees and jimmy for a while. tldr. we are launching strawberry. we wanted to generate some hype. we did.

please burn after reading.

[achilles @killthebuddha_]

what are the chances that @iruletheworldmo can solve today's daily @arcprize puzzle https://t.co/Tb7SRoksTI

i'm not sure i agree with a lot of the ARC creator's more fundamental takes, but ARC is a GREAT way to hype check new models

[achilles @killthebuddha_]

@flydotio I really, actually appreciate the self-aware humor, but this kind of sloppiness is exactly the reason each of my attempts to use fly have failed. When you’re working with outdated docs or silently-failing builds, this sort of branding looks like a red flag.

[Shane Mac @ShaneMac]

Big day here...

Ephemera has acquired @converseapp_, and as of 9 am EST, we open-sourced it.

Converse is the simplest and fastest app built on XMTP. The Converse team has been building with XMTP since day one, and with the ability to reach over 2m+ identities on XMTP, we are excited to put more resources into building a better future together.

We need to make using crypto as easy as sending a message. We have already started implementing secure groups into Converse and we have so much more work to do so everyone in the world has a secure messaging application that is native to crypto.

Everything starts with trust.

The world needs an open, secure, and decentralized communication platform. One that everyone can trust to build on and use.

That’s why we are open-sourcing the application layer, which is also built on the open & secure XMTP network.

Let’s talk about why we believe in open-sourcing everything. Converse and XMTP.

We are building the most open and secure developer platform in the world—one that all developers can trust to build on, unlike the centralized platforms of the past, such as Facebook, WhatsApp, Telegram, Zynga, etc.

I spent the past 15 years of my career building on top of the Big Tech platforms. This means I spent the last 15 years of my career getting the rules changed, promises broken, and ultimately getting my businesses killed or forced to sell because I built on top of centralized platforms.

To protect global communication, we must create a more open and secure communication layer for the internet. We cannot risk building on top of protocols owned by a single company or, in some cases, a single person.

We deserve a messaging network that is more aligned with the entire crypto community and our core values of openness, security, and decentralization.

This isn’t a move about acquiring Converse; it’s about building trust with the best developers in the world and aligning XMTP with the values of the entire crypto community.

It’s a move that makes the application layer as open as the protocol layer. Developers don’t just build on protocols; they build on apps, too. Building on closed-source apps is like playing a game without knowing the rules.

But it doesn’t have to be this way.

Here’s our not-so-secret plan for the next decade which has been the same since day one:

1. Build the most open, secure, and decentralized messaging protocol in the world (v3 upgrade to XMTP brought world-class privacy and security to our protocol)

2. Build amazing and secure messaging applications that are native to crypto - Converse will push the limits to show what’s possible and be open so all developers can benefit from it

3. Open source everything to create the most trusted developer ecosystem in the world

Join us, and let’s build:

We are working hard on an #AllNewConverse and we will be letting people in on an invite only basis starting today!.

This will include secure group messaging and on-chain native functionality.

If you want to try it, shoot me a DM on XMTP @ https://t.co/k8jvcf4c0R

To build with us, check out the Converse or XMTP open-sourced repos below:
https://t.co/3c0JoqHGYC
https://t.co/Djav3EBDvl