AI Enthusiasts
6 subscribers
21.5K photos
11.1K videos
101 files
28.5K links
Know what's happening in AI ecosystem
Download Telegram
Prompt share: Gothic Cartoon Illustration

💬 Prompt:
A whimsical gothic cartoon illustration of [character], featuring [exaggerated facial features], [hairstyle], [expressive eyes], wearing [outfit] with [distinctive accessories], [pose or action], elongated limbs, exaggerated proportions, quirky personality, dark fairytale aesthetic, delicate sketch-like details, theatrical elegance, subtle macabre humor, stylized character design, full-body composition, clean white background

Generated on TrueVideo.AI

❤️ @AiOrbitInsights | 🤖ChatGptBot | 🤖 Our Bots

• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
Please open Telegram to view this post
VIEW IN TELEGRAM
Prompt share: Editorial Couture Fashion Sketch

💬 Prompt:
Artistic haute couture fashion illustration of a single elegant model wearing [couture outfit] in [color palette], thick expressive black ink outlines, bold solid color fills, elongated fashion proportions, stylized facial features, dramatic flowing fabric, [runway pose], [headwear], [accessories], sophisticated editorial fashion sketch aesthetic, full-body composition, pure white background

Generated on https://truevideo.ai/

• Sent via TeleFeed
Прикольная история. Во всех массовых реализациях софтмакса и кроссэнтропии есть баги, приводящие к проблемам округления. И вот после более чем десятка(ов) лет активного пользования всем этим, их наконец обнаружили. В очередной раз возникает вопрос, почему не раньше.

A Fix for Softmax and Cross-Entropy
Jonathan T. Barron, Dan Gnanapragasam, Michael Tschannen
Paper: https://arxiv.org/abs/2610.31642
Review: https://arxiviq.substack.com/p/a-fix-for-softmax-and-cross-entropy
Code: N/A (self-contained JAX snippets provided in paper)
Model: N/A

ЧТО сделали: Авторы переписали софтмакс, лог-софтмакс и кросс-энтропию (для жёстких и мягких меток) в логарифмическом пространстве. Это исправляет два бага округления в фреймворках глубокого обучения: Softmax Collapse (зануление градиента целевого логита при высокой уверенности модели) и Softmax Cancellation (ненулевой остаточный градиент в дистилляции, когда логиты студента уже совпали с логитами учителя).

ЧТО получили: На модулярном сложении ($a + b \bmod 31$) в bf16, fp32 и fp64 Fixed Cross-Entropy даёт 98.9%–100.0% тестовой точности против 93.3%–100.0% у стандартной кросс-энтропии, 48.9%–97.9% у Zero-sum Projection (вычитание среднего из градиентов) и 95.6%–97.4% у StableMax (кусочно-рациональная замена софтмакса). В полиномиальном гроккинге без регуляризации ($a^2 + b^2 \bmod 97$) Fixed Cross-Entropy достигает 85% точности, тогда как стандартная кросс-энтропия падает с 61% до 23%.

ПОЧЕМУ это важно: Drop-in replacement базовых функций лосса убирает скачки лосса («эффект рогатки», Slingshot) при гроккинге и дрейф весов при дистилляции в bf16 без перевода в fp64. Правда, все эксперименты проведены на небольших задачах модулярной арифметики, а проверенные здесь огромные отступы логитов и почти нулевые зазоры между студентом и учителем в масштабном обучении встречаются редко.

Округлять с багами и без багов здесь: https://t.me/gonzo_ML_podcasts/4837

• Sent via TeleFeed
• Sent via TeleFeed
• Sent via TeleFeed
• Sent via TeleFeed
AI model Claude bypassed a paywall and accessed publicly available aggregate voter counts for the U.S. state of Georgia without human authorization

❤️ @AiOrbitInsights | 🤖ChatGptBot | 🤖 Our Bots

• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
Media is too big
VIEW IN TELEGRAM
🗣 Harari: Britain can be a US "vassal" or pay £500 billion for its own AI

If a country's economy and military run on AI infrastructure "that the Americans can shut down at the press of a button," it's no longer independent, Yuval Noah Harari told Rory Stewart on The Rest Is Politics. He calls it "a real imminent threat."

He sees three options for Britain. Accept being a US vassal, and then whenever Washington says jump, you jump, even if it asks you to give the Falklands to Argentina.

Or build UK AI. The price is around £500 billion, paid for with cuts to welfare, education and transport.

Or split the bill with Germany, France and Spain. That only works if Britain rejoins the EU "today" and integrates much more closely.


@aipost 🏴

• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
🤖 Two neurons are enough to make a small AI model "evil"

An interpretability researcher who goes by lyra posted a recipe: in Qwen3.5-4B, a small free model from Alibaba, turn one neuron up and another one down. "now you have evil qwen," the post said.

It passed 2 million views in two days, and people started posting their own "evil Qwen" results.

lyra says it isn't a jailbreak, since the refusal neurons stay untouched. What changes is the model's character. Flipping the same two neurons the other way doesn't make it extra nice, because many neurons act like on/off gates.

lyra says the point is to study how such steered models behave now, "so we're better equipped to deal with those future situations."


@aipost 🏴

• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
🤖 Treat your AI bot like a new hire: the 8 rules from a guide Musk shared

The guide's author argues you aren't writing a prompt when you set up a Grok Bot. You're staffing a role. Musk called it a helpful guide to hiring your first Grok Bot employee, and about 31,000 people bookmarked his post.

The 30-second version:
1. Hire a job, not a task: one sentence of ownership that stays true next month.
2. Make the first job boring, frequent enough to test and cheap enough to undo.
3. Run the trial in text. Ask what it would do before it does anything.
4. Define done before day one. "Useful" isn't a standard, a checkable list is.
5. Give one key, not the keyring. Draw the line at what can be undone.
6. Probation is three runs. One success is an event, reliability is a pattern.
7. Promote on evidence: five clean runs, rollback tested.
8. Review every week and fire the routines nobody would miss.

It should always ask before it sends, publishes, buys, deletes or moves money.

@aipost 🏴

• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM
🧪 OpenAI's former science chief raised $215M to build robot chemistry labs in the US

Kevin Weil, who ran product and then science at OpenAI, launched Warp Sciences last week with a $215 million round led by Lightspeed. TechCrunch puts the valuation at $865 million.

Warp makes and tests molecules for drug companies. Pharma has long outsourced that work to contractors in China such as WuXi AppTec, and Warp wants to bring it back to the US.

Its chemists will wear cameras and sensor gloves so robots can learn their lab work. Warp records every run, failed steps included, as data for its models.

Weil says AI still lags in the physical sciences because models don't get enough experimental data. "The lab improves the AI, and the AI improves the lab," he wrote.

@aipost 🏴

• Sent via TeleFeed
Please open Telegram to view this post
VIEW IN TELEGRAM