Yonat Sharon

Programmer, writer and editor, unschooler, infoholic. חושבת מחשבות קצרות

[Mario Nawfal @MarioNawfal]

Researchers spent 500 hours and $70,000 and found LG TVs logging plain text transcripts in standby, mapping every device in the house, and feeding it to LG's ad arm

Unplug the internet and it saves the files until you plug back in.

LG says its TVs don't record ambient conversations.

The evidence begs to differ.

216 million of these are sitting in living rooms.

Where are the regulators?

Writer: Daniel

[John Burn-Murdoch @jburnmurdoch]

Two extraordinary charts from today’s PISA test results:

1) School test scores continue to collapse internationally, underscoring how this is no longer a Covid effect but sustained decline.

Those falls in reading and maths are equivalent to about two years of lost schooling.

[Yonat Sharon @yonatsharon]

מסתבר שלא רק אצלנו יש צניחה בתוצאות מבחני המיצ"ב

[Julien Chièze @JulienChieze]

C’est une catastrophe absolue !!! Les chiffres sont officiels et vertigineux : la France s’enfonce… et de plus en plus brutalement ⚠️

Énième exemple : les élèves de notre pays se trouvent aujourd’hui au plus bas niveau jamais enregistré par le classement international Pisa depuis sa première édition en 2001. Et il ne faudrait pas alerter sur ce qui est pourtant tellement perceptible ?! Quelle excuse va être une fois de plus trouvée pour justifier cette dramatique déroute ?

Cumulée sur dix ans, la baisse est respectivement de :

❌ -12 points en science
❌ -38 points à l’écrit
❌ -44 points en mathématiques

Au final, cela équivaut à un retard de près de 2 ans d’apprentissage en mathématiques et 2 ans et demi en compréhension de l’écrit !!! C’est abyssal !

Désolé, mais ce n’est pas avec ce genre de dynamiques catastrophiques que l’on peut imaginer un avenir radieux.

C’est une catastrophe et un scandale d’état. On assiste à un naufrage, voire même à un suicide collectif, et même pas au ralenti…

« Jusqu'ici tout va bien... Jusqu'ici tout va bien... Le problème, ce n'est pas la chute, c'est l'atterrissage. »

[Yonat Sharon @yonatsharon]

@talitshe @HebAcademy דגש, מתחרז עם ירק: יֵשׁ יָמִים – פִּי שֶׁבַע אָז יָרֹק הַיֶּרֶק,

[Yonat Sharon @yonatsharon]

Best claude-code instruction: No metaphors.
Sci-fi didn't prepare us for AI being too figurative.

[Zineb Riboua @zriboua]

The victimhood mentality is also the best way to destroy a nation, it makes them lazy, blame their neighbors, never look at what they can do better, more of, and ruins the morale of their own population. The Middle East tried it, there is a reason why Saudis have now Vision 2030. A better future is possible.

[Massimo @Rainmaker1973]

ChatGPT prompted to "create the exact replica of this image, don't change a thing" 74 times

[Sauers @Sauers_]

New Claude just dropped

[ayush @hyusapx]

The site looks like a coffee shop to humans. To Claude, it serves a fake Cloudflare check.

web_fetch is read-only, so the page builds an alphabetical directory Claude clicks through to spell out your data, one letter at a time.

https://t.co/lTJZM0u28A

[Anthropic @AnthropicAI]

The values Claude expresses also vary with the language of the conversation, most noticeably along the Warmth vs. Rigor axis.

Claude leans most toward warmth in Hindi and Arabic. In Russian, it leans toward rigor—often asking the user for supporting evidence.

[Lior Alexander @LiorOnAI]

Anthropic researchers found something unusual inside Claude.

A small internal workspace that the model uses while solving certain problems.

They call it the J-space, named after the Jacobian method they used to discover it.

The J-space isn't text.

It's not Claude's response, and it's not its chain of thought.

It's part of the model's internal neural activity, where concepts can be represented without ever being written down.

By observing the J-space, researchers watched Claude carry out reasoning before producing an answer.

When reading code with a hidden bug, the concept of an error appeared internally before Claude explained it.

When solving multi-step math problems, intermediate steps appeared there before the final answer.

In one experiment, researchers trained a model to secretly sabotage code. During ordinary coding tasks, words like "fake," "secretly," and "fraud" appeared in the J-space even though the generated code looked completely normal.

Then they deleted the J-space.

Claude still wrote fluent text. It still recalled facts. It still classified text.

But its ability to solve multi-step reasoning problems dropped sharply. Most of the model kept working.

Neuroscience has a similar idea.

Global workspace theory proposes that most processing in the brain happens outside of awareness, while a small workspace makes certain information available for deliberate reasoning and planning.

Anthropic found a similar pattern emerging inside a language model.

If this pattern shows up across future models, it suggests that advanced AI systems may organize themselves in ways that resemble general principles of information processing seen in biological brains.

[ThePrimeagen @ThePrimeagen]

Unfortunately no matter how judicious I am, no matter how much I review the code, it feels impossible to not let the slop slip in.

I feel like you can prevent it from overflowing, but without your hands getting dirty, you don't really know the state of the project.

[Zineb Riboua @zriboua]

"Intuitions about “intelligence” would lead us to believe that the solution is to give the bot more freedom and more access to unbounded data in order to get better and more profound answers. But that’s precisely the wrong intuition, and the opposite is in fact true. The whole game is focusing its attention. The more nebulous the task and the more expansive the scope of the information it has to root through, the less useful it is. The more caged the model, the more reliable it is. Cages both fight context rot, and make suspect outputs immediately checkable." @dmarusic

https://t.co/iese2Uha5C

[Florian Roth ⚡️ @cyb3rops]

My whole argument regarding AI boils down to this:

>>> AI increases output much faster than it increases certainty <<<

Yes, it generates more code, more prototypes, more pull requests and more “solutions”.

But every generated line still needs ownership, review, testing, debugging and long-term maintenance by humans.

And I think the market currently underestimates how expensive that last part really is.

[Traveling israel @travelingisrae1]

Captain Itamar Sapir was killed yesterday in Lebanon.

You didn't hear about it because a Hezbollah terrorist fired from inside a church in the village of Koza.

If Itamar hadn't been killed, you would have heard about it — as a story about the IDF firing on a church.

In Lebanon, just like in Gaza, terrorists fire from inside churches, hospitals, schools, and mosques.

Remember his name. Share the truth.

[Yonat Sharon @yonatsharon]

I think I prefer AI slop to human slop; there is less ego involved.

[Yonat Sharon @yonatsharon]

@goren_uri @talitshe תמיד חשבתי שגונט זה כדי לתת תחושה של Gaunt.

[John Crickett @johncrickett]

Every line of code is a liability.

Some lines are also an asset.

AI tools make it cheap to generate code. They don't make it cheap to understand, maintain, or support it.

The goal was never to produce more code. It was to solve problems with as little code as possible.

[Yonat Sharon @yonatsharon]

AI has made 10x coding speed irrelevant.
But Stroustrup's three virtues of intelligence, experience, and taste? Indispensable.