BuddZ

How disparate concepts related to each other. The layers thesis, one filing at a time.

[BuddZ @buddzzh]

Eight days. Five bond deals. One graph.
The shortest path from a chat window to a credit default swap is four edges.
New: Four Hops to a Default Swap
https://t.co/PwVKX8inrz

[BuddZ @buddzzh]

Using X from 8 years

Still didn’t figure out it can make you meet a lot of incredible person

When I want to search some news it is my first location where to look

Ah, you can get also a screenshot of a post using the copy button 🤝

[BuddZ @buddzzh]

@teneo_protocol @omarsar0 Only by having the agent keeping the graph update means the agent must autonomously decide when and if using tools

[BuddZ @buddzzh]

@neo4j Looks promise. Let your agents build its own memories in a KG via an optimized neo4js engine

[BuddZ @buddzzh]

@RapidataAI @GoogleDeepMind @NewsFromGoogle @DynamicWebPaige @OfficialLoganK

[BuddZ @buddzzh]

This is not a “small model vs big model” story.
This is the first clear signal that the cost curve just bent harder than most people are pricing.
Full breakdown of the numbers (HLE-Diamond, Vals, Terminal-Bench, legal/tax splits) in the next posts.

[BuddZ @buddzzh]

Stop scrolling.
Meta just released a model that matches DeepSeek’s performance
at 1/4 the cost
and they refuse to publish the parameter count or architecture.
Ember sits inside the error bars of a 552B sparse MoE
while remaining the cheapest model on the entire agentic index.

[BuddZ @buddzzh]

If interested to such type of articles Interact with the post as much you can

[BuddZ @buddzzh]

Stop scrolling

That’s the magic every one is looking for

We’ve investigated how the magic persuasion of the social algorithms manipulate your brain becoming part of your most hidden thoughts

+1.5k articles, blogs and real life stories collected

What are they saying ?

[BuddZ @buddzzh]

On HLE-Diamond, Artificial Analysis has them tied: DeepSeek 25.3% (34.8 reasoning, 15.8 knowledge), Ember 25.1% (36.6, 13.6).

A 552B sparse MoE and Meta's smallest model land in the same place. Ember wins reasoning by 1.8; DeepSeek wins knowledge by 2.2.

[BuddZ @buddzzh]

Meta's Muse Ember is rhe cheapest model on Val's agentic index: 60.48% at $0.15 per test, 15 min.

Ember is 0.78 behind Deepseek V4.1 Flash - inside the error bars - at a quarter od Deepseek's cost. 1.75 ahead of Haiku 5.5 at 1/21 the cost. 8.16 ahead of GPT-6 Luna at a third.

[BuddZ @buddzzh]

On Vals' components, it's not "DeepSeek is smarter."

Ember wins legal (75.69 vs 71.86), Harvey's legal agent benchmark (25.0 vs 14.2), and Terminal-Bench 4.0 (39.4 vs 31.8). DeepSeek wins tax (44.4 vs 38.2) and Vibe Code (87.7 vs 83.7). Finance ties: 60.91 vs 60.89.

[BuddZ @buddzzh]

The AI credit web is not a chain. It is a set of cycles.

OpenAI→Oracle $300bn. Microsoft→OpenAI $138bn vs OpenAI→Microsoft $250bn. CoreWeave↔Nvidia.

The gross arrows are ~$1.9tn. The net delta is the missing number.

https://t.co/PwVKX8hPC1

[Bloomberg @business]

CoreWeave sent a letter to Texas Governor-elect Gina Hinojosa seeking to work with her incoming administration on present and future data center projects in the state https://t.co/CurIfjlj1R

[BuddZ @buddzzh]

HLE-Diamond's leaderboard lists nine models. TwIL-8B is not one of them.

That is the other half of fresh: the test is new enough, and expensive enough, that the small model is often not in the room.

Full audit: https://t.co/bDkitfUKgM

[BuddZ @buddzzh]

Small models look caught up on the benchmarks that stopped moving. Not the new ones.

TwIL-8B's model card: 98.2 on MATH-500, 12.6 on Humanity's Last Exam. Frontier: 45.8–48.7.

Same 8.8B model. One test is saturated. One is not.

[BuddZ @buddzzh]

HLE-Diamond (Sept 2026) splits 1,000 questions into 500 reasoning, 500 knowledge.

Best model on it, no tools: 74.2% reasoning, 45.6% knowledge.

The wall is the knowledge half: it scales with parameters more than with post-training.

[BuddZ @buddzzh]

@vineerpasam Did you fine tuned with a large learning rate ?