Pinned Tweet
We present DyPE, a framework for ultra high resolution image generation.
DyPE adjusts positional embeddings to evolve dynamically with the spectral progression of diffusion.
This lets pre-trained DiTs create images with 16M+ pixels without retraining or extra inference cost.
🧵👇
Guy Yariv retweeted
We found an interesting phenomenon: starting a CoT with a self-instruction makes it significantly more controllable than giving the same instruction in the prompt. Interestingly, it also helps evade CoT monitors.
Guy Yariv retweeted
Wow, Fibo 2.0 ranked #1 *open weights* model in image editing 🎉
Still JSON-native control
Turbo for fast inference, Dev for fine-tuning
High-quality general editing
Better text, texture & anatomy
>>>
Guy Yariv retweeted
Today we’re introducing Base Code.
It’s an early preview of how we think software will be built in the future.
It’s a product that encapsulates everything we’ve learnt from:
- How our users are building software.
- How we’re building internally, scaling to hundreds of millions of dollars in revenue while keeping our engineering team very small and focused.
Here are some of the principles behind it, which align with how we at @Base44 think about the future of software engineering:
Cloud:
- Software will move past local desktop environments (where current tools are widely used) and move to the cloud.
- Moving everything to the cloud is not easy. Setting up dev environments with databases, infra components, services, mock data, etc. is easier said than done.
- Base Code first scans your code repo and sets up everything - every infra component (databases, Redis, etc.) - in the cloud. It runs nonstop until your preview environment is ready.
- From an enterprise standpoint, it’s also the logical thing to do: a centralized, governed dev environment instead of handing out keys and secrets to all team members.
- Once Base Code does that, EVERY team member can work in this environment. No more syncing local environments.
Internally, this (cloud) is one of the main things that enabled us to move so fast.
Collaboration:
- Once everything is in the cloud, everybody can write software from anywhere (any browser, your phone, WhatsApp / iMessage coming soon).
- You can easily see what everyone is working on, where they’re at, and how they’re prompting - and jump to their environment in one click.
- We’ve built many great collaboration features from the ground up.
Loops, automations, software factories:
- A full, working cloud environment allows for many advanced capabilities we will unveil soon-think agents running in the browser and testing on every device and in any browser.
- It also allows the use of automations to build software factories-e.g., agents reading support tickets, identifying bugs to fix or features to develop, implementing the changes, verifying them in the cloud, and potentially pushing them.
And lastly, there’s a real advantage in being model agnostic.
--------------
As always, we’re releasing it very early. It’s far from perfect-but we’re looking for early feedback so we can build it together with our great community and make this vision a reality.
*We’re giving it away for free for 30 days*.
Give Base Code a try. Connect any GitHub repo, get your live preview, and share it with your team.
I’d love to hear what works and what doesn’t:
app.base44.com/base-code
🤖 Made with AI
Guy Yariv retweeted
I'm super bored of LLM-stylized papers/blogs. I think we all are?
So I asked Claude to explain its tells: web.mit.edu/phillipi/www/cla…
Let's avoid these! I'd much rather read *your* words, see *your* rickety figures, than face down yet another of Claude's anodized corporate veneers.
Guy Yariv retweeted
Lesson from the DeepSeek-V4.1 technical report:
"...the marginal return of engineering the data and environment pipeline substantially exceeds that of algorithmic novelty in post-training."
My 2 cents. The interesting part isn't data > algorithms. It's that the distinction is collapsing. Look at what their pipeline actually does: synthesizes verifiable tasks with reward signals, builds interactive agent environments, calibrates difficulty and curriculum at scale. That's not just data collection. That's research. Some of the highest leverage research is happening in the data + environment pipeline.
We are in the era of research AND scaling up.
The current recipe doesn't need to work forever. It only needs to help us find the next S curve.
Guy Yariv retweeted
A big part of my work these past few months has been dedicated to working on Base 1.
Proud of this moment and the team that worked on it and even more excited for all of the road ahead!
Announcing Base 1’s benchmarks:
After rigorous testing in production, attached below Base 1 results for building beautiful, functional and complex web apps inside @Base44.
we released Base 1, our own model, a few weeks ago. we've been measuring it since.
For app building, there are no standard benchmarks. We A/B test in production with millions of builders, who we believe provide the best benchmark for measuring real value across real-world use cases.
Base 1 came close to Fable, while being 4.5x cheaper.
Base 1 has the lowest frustration of anything we've tested. below Sonnet 5, below Opus 4.8, below Fable 5.
Fable 5 still edges it on success rate, 76.4% to 74.9%.
i'll post the full benchmarks in the comments
Guy Yariv retweeted
Google Deepmind argues that LLMs can never make real scientific discoveries.
They published a paper breaking down Albert Einstein’s private view of scientific discovery.
In a famous letter to his friend Maurice Solovine, Einstein drew a diagram of how science actually happens.
It is a cyclical loop.
First, you experience raw sensory data. Then, through a mysterious, non-logical act of intuition, you make an intuitive "jump" to abstract axioms. Finally, you use strict logical deduction to derive consequences from those axioms.
Generative AI has completely mastered two-thirds of this loop.
• Induction: Statistical pattern matching across billions of tokens.
• Deduction: Formal proof generation, like AlphaProof solving complex math Olympiads.
AI can crunch data and it can prove theorems.
But it cannot make the jump.
The paper argues that AI completely lacks Abduction, the generation of novel explanatory hypotheses when observational data is scarce.
The prevailing tech myth says that "creativity is just data compression." That if you feed an LLM enough text, scientific breakthroughs will naturally pop out.
Einstein’s formulation of General Relativity proves that is a delusion.
When Einstein formulated relativity, the observational data didn't demand a new physics framework; classical mechanics was still massively successful. The breakthrough required a conceptual rupture. An intuitive leap from physical reality to a brand-new set of foundational axioms.
An LLM can execute the math once the axioms are given. But it is structurally incapable of formulating those premises on its own.
It can interpolate inside existing human thought, but it cannot transcend it.
The translation of physical reality into formal axioms remains the absolute, hard bottleneck of artificial scientific invention.
We can build models with trillions of parameters. We can scale compute into the stratosphere.
We can make the calculator infinitely fast.
But until we solve grounding, the machine can process all the data in the universe.
It still can't make the jump.
We’re presenting DyPE at #ICML2026 tomorrow, so in case you’re there, come check out our poster:
Wed, Jul 8, 14:30 KST
Poster Session 4
If you’re at #ACL2026, check out my oral presentation on augmenting LLMs via late multi-image fusion
July 6, 11:00–12:30
Harbor B-C
Our paper:
"LaMI: Augmenting Large Language Models via Late Multi-Image Fusion"
has been selected for an Oral Presentation at #ACL2026!
LaMI boosts LLM visual commonsense by generating complementary images from a text prompt and late-fusing their evidence into the prediction
🧵
If you unfortunately miss the oral presentation tomorrow, the recorded presentation of LaMI is now on YouTube:
youtu.be/Sb7LkYQUI5A
98
Our paper:
"LaMI: Augmenting Large Language Models via Late Multi-Image Fusion"
has been selected for an Oral Presentation at #ACL2026!
LaMI boosts LLM visual commonsense by generating complementary images from a text prompt and late-fusing their evidence into the prediction
🧵
Recorded presentation of LaMI is now on YouTube:
youtu.be/Sb7LkYQUI5A
62
Guy Yariv retweeted
I’m at #ACL2026 this week! Presenting StressTest on July 5th, Poster Session A (11:00–12:30, Coronado).
If you’re into Speech LM reasoning and multimodal grounding feel free to DM me if you'd like to grab a coffee or chat! ✌️🗣️
🔗 Project page - pages.cs.huji.ac.il/adiyoss-…
Guy Yariv retweeted
1/7
New paper 🧵 ScenA generates a multi-speaker audio scene: overlapping speech, laughter, real room noise, from a text description and a few reference voices.
When trained in the obvious way, it ignores the text and decides who speaks on its own. We found out why and fixed it.
Guy Yariv retweeted
1/6 Diffusion models are scaling up, but deploying a massive, monolithic network uniformly across the entire generative timeline is inherently inefficient.
Introducing Complexity-Balanced Splitting (CBS): a principled framework that allocates capacity exactly where needed!👇🧵
Guy Yariv retweeted
📢🚀 Happy to share our recent work.
SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers
📄Paper: arxiv.org/abs/2605.22668
🌐Webpage: rajabi2001.github.io/sega/
🤗HF: huggingface.co/papers/2605.2…
Guy Yariv retweeted
🚀New Paper🚀 Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale
AI agents are increasingly deployed across critical environments. How do we know they're solving tasks the intended way, and not just appearing successful by hacking the reward?
Details👇🏽
Details, paper, and code:
Project page: guyyariv.github.io/LaMI
arXiv: arxiv.org/abs/2406.13621v2
Code: github.com/guyyariv/LaMI/
1
4
585





























