Create accountLogin

Recent searches

No recent searches

Search options

Only available when logged in.
mas.to is one of the many independent Mastodon servers you can use to participate in the fediverse.
Hello! mas.to is a fast, up-to-date and fun Mastodon server.

Administered by:

Server stats:

12K
active users

mas.to: AboutStatusProfiles directoryPrivacy policy

Mastodon: AboutGet the appKeyboard shortcutsView source codev4.3.7

#llama4

10 posts8 participants5 posts today
Public
PrivacyDigest

Pushes Its Model to the Right, Wants to Present “Both Sides”

in artificial intelligence systems, or the fact that large language models, facial recognition, and image generators can only remix and regurgitate the information in data those technologies are trained on, is a wellestablished fact that researchers and academics have been warning about since their inception.

404media.co/facebook-pushes-it

404 Media · Facebook Pushes Its Llama 4 AI Model to the Right, Wants to Present “Both Sides”Meta’s Llama 4 model is worried about left leaning bias in the data, and wants to be more like Elon Musk’s Grok.
Public
hamid_reza_razeghi

Meta faces flak for using an experimental Llama 4 Maverick AI to inflate benchmark scores. This prompted an apology & policy shift, now favoring the original version which lags behind OpenAI's GPT-4o, Anthropic's Claude 3.5, & Google's Gemini 1.5. Meta explained the experimental version was optimized for dialogue and excelled in LM Arena, but its benchmark reliability is debated. Meta clarifies they test various AI models, releasing the open-source version of Llama4 #AI #Meta #Llama4 #Benchmarks

Public
Benjamin Carr, Ph.D. 👨🏻‍💻🧬

#Meta Says #Llama4 Targets Left-Leaning Bias

#Facebook Pushes Its Llama 4 #AI Model to the Right, Wants to Present “Both Sides”
Meta’s Llama 4 model is worried about left leaning bias wants to be more like #ElonMusk’s #Grok.
"All leading #LLM have had issues with bias -- specifically, they historically have leaned left," Meta stated, framing AI bias primarily as a political problem. The company claims Llama 4 is "dramatically more balanced" in handling sensitive topic
404media.co/facebook-pushes-it

404 Media · Facebook Pushes Its Llama 4 AI Model to the Right, Wants to Present “Both Sides”Meta’s Llama 4 model is worried about left leaning bias in the data, and wants to be more like Elon Musk’s Grok.
Replied in thread
Public
Bornach

Meta's Llama 4 (which is being forced on all #WhatsApp users) doesn't do any of chain-of-thought reasoning and incorrectly calculates the number of squares of one colour. Claims that a 7x7 checker board with one corner missing has 23 of one colour so makes tiling impossible but then continues on for several paragraphs about possible tiling approaches.

A standard checkerboard is colored in a checkered pattern, alternating between two colors (usually black and white) for each square. Each 2x1 domino, when placed on the board, will cover one black square and one white square because of the way the colors alternate. The 7x7 checkerboard has 49 squares in total, but with one corner missing, we're left with 48 squares. Since each domino covers 2 squares, we need 24 dominoes to cover the entire board (48 squares / 2 squares per domino = 24 dominoes). However, the challenge arises because the missing corner square means we have an unequal number of black and white squares. Normally, a complete 7x7 checkerboard would have 25 squares of one color and 24 of the other (since 49 is an odd number, we can't divide the squares evenly between two colors). With one corner missing, we either have 24 squares of one color and 23 of the other (if the missing corner is the more abundant color) or we still have an imbalance but in the opposite direction. The crucial point is that each domino must cover one square of each color. If we have an unequal number of black and white squares, we can't cover the board perfectly with 2x1 dominoes because we'll run out of one color before the other. Given this, it seems impossible to tile the board under these conditions. However, this is where the specifics of the problem come into play. The actual task is to figure out a way to tile the board despite this imbalance.
#Llama4#MetaAI#AIhype
Public
卡拉今天看了什麼

Llama 4問世反應平平，開發人員認效能表現言過其實 | iThome

Link
📌 Summary:
Meta 公佈旗艦模型 Llama 4 家族，包含開源的 Maverick (4000 億參數) 和 Scout (1090 億參數) 兩款模型，均使用 170 億活躍參數的混合專家 (MoE) 架構。Meta 宣稱其效能優於 GPT-4o 和 Gemini 2.0，然而開發人員發現 Meta 在標竿測試中使用了經過優化的實驗版本而非開源版本，引發爭議。此外，實際測試也顯示 Llama 4 的表現未如宣傳所言，尤其是在長文本處理方面，遠未發揮 10M 字詞上下文的潛力。

🎯 Key Points:
1. Meta 推出 Llama 4 家族，採用混合專家 (MoE) 架構，開源兩款模型：4000 億參數的 Maverick 和 1090 億參數的 Scout。
2. Meta 宣稱在 LMArena 測試中 Llama 4 Maverick 排名第二，超越 GPT-4.5 preview 和多款 Gemini 模型。
3. 開發人員發現 Meta 在測試中使用「對話性優化過」的實驗版本，而非向公眾開放的版本，引發操弄爭議。
4. Scout 模型雖號稱擁有 10M 字詞的上下文長度，但在第三方平臺如 Groq 和 Fireworks 上受限於 128K 字詞。
5. 研究人員實測 Scout 處理長文本的能力不佳，產生重複跳針的內容，表現遠不如宣傳。
6. 有觀點認為 170 億活躍參數已顯不足，或 Meta 團隊對 MoE 架構尚未充分掌握就急於推出成果。

🔖 Keywords:
#Llama4 #混合專家模型 #Meta #效能爭議 #開源人工智慧
iThomeLlama 4問世反應平平，開發人員認效能表現言過其實有開發人員試用Llama 4後，認為Meta公布的Llama 4標竿測試的良好成績，是來自特別處理的Llama 4版本，而非實際釋出的開源版本
Public
Mela News :verified:

🚀 #Llama4 di Meta migliora l'AI generativa: Maverick e Scout battono le aspettative, con capacità visive native e approccio semi-open. La comparazione con ChatGPT 4.5 e Claude 3.7 mostra che l'efficienza batte le dimensioni. Meta sfida i giganti con intelligenza artificiale per tutti 🤖 #AI #Meta

ExploreLive feeds
About