Evyatar Bluzer

[Boy George @BoyGeorge]

In Buddhism we understand that every action comes with a consequence. Due to my decision to stand up for my Jewish friends I am losing a few people from my life. Today I parted company with Tony Pontius who has run my record label BGP for a number of years. I wanted to release 'We Will Dance Again' but he was unequivocal and said 'I want nothing to do with this song' and I responded 'ok, we are done'.
When someone is gutless it really does not hurt to say goodbye. My faith in humanity will not diminish and the truth will always win.

I will release it!

[Anthropic @AnthropicAI]

New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel.

Here's what it taught us about the future of autonomous software development.

Read more: https://t.co/htX0wl4wIf

[Evyatar Bluzer @Evyatar_Bluzer]

Papers (cont.):
5. Test-Time Compute (arXiv:2408.03314) - Scale models used by uncertainty. Gap: Tier-based, not per-query optimization.

[Evyatar Bluzer @Evyatar_Bluzer]

I was tired of consulting with multiple LLMs.

So I vibe-coded this side project - a tool to automatically compare answers from multiple AI models for questions I really care about (think: health, finance, kids' education).

I was humbled when @karpathy shared a similar idea :)

[Andrej Karpathy @karpathy]

As a fun Saturday vibe code project and following up on this tweet earlier, I hacked up an **llm-council** web app. It looks exactly like ChatGPT except each user query is 1) dispatched to multiple models on your council using OpenRouter, e.g. currently:

"openai/gpt-5.1",
"google/gemini-3-pro-preview",
"anthropic/claude-sonnet-4.5",
"x-ai/grok-4",

Then 2) all models get to see each other's (anonymized) responses and they review and rank them, and then 3) a "Chairman LLM" gets all of that as context and produces the final response.

It's interesting to see the results from multiple models side by side on the same query, and even more amusingly, to read through their evaluation and ranking of each other's responses.

Quite often, the models are surprisingly willing to select another LLM's response as superior to their own, making this an interesting model evaluation strategy more generally. For example, reading book chapters together with my LLM Council today, the models consistently praise GPT 5.1 as the best and most insightful model, and consistently select Claude as the worst model, with the other models floating in between. But I'm not 100% convinced this aligns with my own qualitative assessment. For example, qualitatively I find GPT 5.1 a little too wordy and sprawled and Gemini 3 a bit more condensed and processed. Claude is too terse in this domain.

That said, there's probably a whole design space of the data flow of your LLM council. The construction of LLM ensembles seems under-explored.

I pushed the vibe coded app to
https://t.co/EZyOqwXd2k
if others would like to play. ty nano banana pro for fun header image for the repo

[Evyatar Bluzer @Evyatar_Bluzer]

Papers (cont.):
4. DeepConf (arXiv:2508.15260) - Confidence-weighted aggregation by cluster size as proxy. Gap: External semantic analysis only, no internal signals.

[Evyatar Bluzer @Evyatar_Bluzer]

@osanseviero @NanoBanana @NanoBanana create an image of (7.7597460, 98.3043384) at sunset

[Evyatar Bluzer @Evyatar_Bluzer]

@karpathy Awesome, I just posted about a very similar project I vibe coded that uses an LLM counsil to improve overall accuracy/quality. Getting some really good feedback from the early users :)

here: https://t.co/bqj4LSP9pm
and on Linkedin: https://t.co/Bqrqh4l5Oe

[Evyatar Bluzer @Evyatar_Bluzer]

So now we just need to pair this with an agent that creates product ideas and together they will iterative and invent ultimate, irresistible, products 🤑

[Garry Tan @garrytan]

You can just prompt things

[Garry Tan @garrytan]

You can just prompt things

[Evyatar Bluzer @Evyatar_Bluzer]

Samsung’s result reframes the future of LLMs:
Progress won’t just come from scaling, but from intelligent design.
A smaller, reasoning-optimized model can now match, or exceed, massive frontier models in real-world benchmarks.

[Evyatar Bluzer @Evyatar_Bluzer]

Samsung Research has introduced a compact language model that outperforms leading large-scale LLMs - including GPT-4 and Gemini 2.5 Pro - on reasoning benchmarks, despite being orders of magnitude smaller in parameter count.

https://t.co/yz5YYlH6AZ

[Evyatar Bluzer @Evyatar_Bluzer]

It also opens competitive pressure on major mobile AI platforms (Apple Neural Engine, Google Tensor, Qualcomm Hexagon) to support models that maximize on-device reasoning efficiency.