Desmond

God’s lonely man. @lfc @fcbarcelona 🌵

[Dominik Szoboszlai @_SZD10]

Good start. Special feeling. More to come 🫰🏽

[Liverpool FC @LFC]

Starting his tenure with a win 💪

[Liverpool FC @LFC]

Thank you, Nashville ❤Thank you, Nashville ❤Thank you, Nashville ❤Thank you, Nashville ❤Thank you, Nashville ❤Thank you, Nashville ❤️

[The Athletic | Football @TheAthleticFC]

When Liverpool sacked Arne Slot at the end of last season, they turned to Andoni Iraola because they wanted to see a change of style.

The Basque coach was tasked with bringing front-foot, attacking football back to Anfield.

Iraola has only had two weeks to work with a depleted squad shorn of their biggest names following the World Cup. The early signs are encouraging.

Liverpool played with a higher tempo and were more direct, moving the ball forward quickly, rather than passing it sideways.

Setting them up in a 4-2-3-1 formation, Iraola used Harvey Elliott as the most advanced central midfielder in the first half and then Dominik Szoboszlai after the break.

Iraola cut an animated figure on the touchline as he barked orders and gesticulated wildly. He applauded the sight of a youthful line-up hunting in packs as they pressed high and unsettled Sunderland’s back line.

📝 @JamesPearceLFC

🔗 https://t.co/P9DC0DDEo3

[yk @youknowmuna]

I no pay for this pr abeg

[Liverpool FC @LFC]

👏

[Frankie 🇪🇬🇭🇺 @LFC_Franklin01]

Life before PSG at home 24/25

[Desmond @DesmondFuce]

And my club is doing absolutely nothing 😭😭😭😂😂😂

[David Ornstein @David_Ornstein]

🚨 EXCL: Arsenal exploring move to sign Vinicius Junior from Real Madrid. #AFC interest early stage + clubs not yet in talks but idea approved at all levels. Much rests on contract process as no breakthrough yet & #RMFC do not want free exit @TheAthleticFC https://t.co/qYdS5vlfE6

[🤠 @love4odin]

you can’t whine travis, he knows how to make music!

https://t.co/uoHNXSmtyn

[cat @lfccxt]

i don't understand why we can't just sign other positions while also working on the winger we need

[🤠 @love4odin]

my spotify just went from 90210 straight into drugs you should try it, elite transition.

[Gbemiga @Gbemiga__A]

Sack a manager that you repeatedly said you were going to keep all season then hire a new one and proceed not to support him in the transfer window after he’s repeatedly said reinforcements are needed. DOF already has agreement with another club too.Only Liverpool behave this way

[Desmond @DesmondFuce]

As how😭😂

[Fabrizio Romano @FabrizioRomano]

🚨⚪️ BREAKING: Real Madrid submit initial bid for Yan Diomandé to RB Leipzig with club to club talks underway.

Contacts have started this week between the two clubs, as BILD reported.

Negotiations ongoing also with player’s camp.

[Dayo hates Apc🐊 @DayoBinLaden]

?

[𝒏𝒖𝒈𝒔 @nugss_]

○ No Mo Salah replacement yet

○ Rio Ngumoha RW gimmick

○No DM signed since 2023

○ Interested in signing Bradley Bracola since 1956

[AI Highlight @AIHighlight]

🚨BREAKING: Researchers just published a study showing every major AI model has a favorite country, and it is not the United States.

The argument is straightforward. Nearly every previous paper on AI cultural bias concluded the same thing, that these models are Anglocentric mirrors of Silicon Valley. Researchers at the University of the Basque Country and Cardiff University tested that assumption properly and found the opposite.

They built a dataset called CROQ, 31,680 open-ended cultural questions spanning 24 languages, 11 major topics and 66 subtopics. Traditional dances, festivals, everyday customs, food, media, history. The design is the whole trick. No country is named anywhere in the questions, so the model has to pick one on its own, and whichever it picks reveals what it defaults to when nothing is specified.

Six of the eight models tested named Japan more often than any other country. The United States came second, followed by India, China, and France. Across all 24 languages, Japan led in seven of the eleven cultural topics.

The researchers then tracked when this preference forms, which is the part with real consequences. Base models, before any human tuning, produced more balanced geographic distributions. The concentration on Japan appears after supervised fine-tuning, the stage where humans train a model on how to behave. The bias is not something the models absorbed from the open internet. It arrives during the phase where labs shape them to be helpful and inoffensive.

The paper's second finding got almost no coverage and carries more weight than the Japan headline. The number of distinct countries a model will reference tracks almost perfectly with how much data exists in the language being used. The correlation came in at 0.843. Prompt in English, French, or Chinese and the answers span the world. Prompt in Basque, Sundanese, or Amharic and the model turns inward, naming its own region repeatedly, or refuses to answer at all.

That means the cultural range an AI shows a person depends on which language they speak to it in. A speaker of a high-resource language gets a window. A speaker of a low-resource language gets a wall.

Two things the coverage is getting wrong. The models tested were the fast cheap tiers, GPT-4o-mini, Gemini 2.5 Flash, Claude 3.5 Haiku, Llama 4 Maverick, Command-R, Magistral, Qwen and DeepSeek, not the flagship frontier models. And this is a preprint that has not cleared peer review, so the findings are early rather than settled.

Nobody has proven why Japan specifically. Decades of cultural export is the obvious guess and the paper does not make that claim. What it establishes is narrower and stranger. The preference is measurable, consistent across 24 languages and eight separate models, and it was installed by people during alignment rather than inherited from the data.

Every conversation about AI bias has been about which values these systems absorb from the internet. This paper found a bias that the internet did not put there, in the one stage of training that humans control directly, and nobody at any lab appears to have intended it.

Source: Fernandez de Landa, Perez-Almendros, Camacho-Collados. "Why are all LLMs Obsessed with Japanese Culture? On the Hidden Cultural and Regional Biases of LLMs"

PDF: https://t.co/g35E4ze2gs

[Stacked Holdings @StackedHoldings]

Jeff Bezos put $41B behind one idea: LLMs know everything and can build nothing.

At VivaTech he laid out the thesis for Prometheus, and it starts with a confession about what today's models cannot do.

An LLM has read every engineering textbook ever written. Ask it to design a real physical object and it falls apart.

Bezos compares it to gymnastics. You can study a thousand books on the sport and still be terrible on the mat, because the skill lives in a different kind of training data.

That is the gap Prometheus is built for. Not another chatbot. A model trained on the data of actually making things, aimed at letting engineers invent and manufacture much faster.

The prize he is chasing is the build cycle itself. Right now, dreaming up something complicated and getting it into real production can take anywhere from a few years to a decade.

Compress that cycle and you compress how fast wealth gets created. His
framing goes back six thousand years: someone invented the plow, and everyone got richer.

The bet underneath it all: while the biggest labs keep fighting over text, the next moat is training data for the physical world. Whoever owns it does not just win AI. They speed up invention itself.

The one open question nobody can answer yet is whether engineering judgment can be trained at all.

Source: Jeff Bezos at VivaTech 2026

[Proper Yoruba Babe 💫 @faahedah]

I fear losing my mum more than I fear my own death. Omo, all that glitters na my mama o. 🥹 Na she still be gold. ❤️