Dipak kumar Tomar

I am a fullstack developer exploring ai/ml field

[Dipak kumar Tomar @Dipaktomar73]

@davidchie Compute is the new king

[Dipak kumar Tomar @Dipaktomar73]

Source: https://t.co/9PVl5Zcg77

For daily AI updates → follow @Dipaktomar73

[Dipak kumar Tomar @Dipaktomar73]

Norway just banned generative AI in elementary schools.

Kids 6–13 lose access from August. Prime Minister: "The most important thing is that children learn to read, write and do mathematics."

Smartphones already banned. More funding for printed books.

Is this the right call?

[Dipak kumar Tomar @Dipaktomar73]

NVIDIA just swept all seven MLPerf Training 6.0 benchmarks.

GB300 NVL72 trains up to 1.6× faster.

DeepSeek-V3 671B: 2.02 min. Llama 3.1 405B: 7.07 min.

8,192 GPUs. Scaling is the new frontier.

https://t.co/aQZ9qAoWqs

[Dipak kumar Tomar @Dipaktomar73]

Source: https://t.co/LJkdxe4D1u

For daily AI updates -> follow @Dipaktomar73

[Dipak kumar Tomar @Dipaktomar73]

MiniMax just launched a frontier model that costs $1 per hour to run.

M2.5 scores 80.2% on SWE-Bench — at 1/10th the price of Opus, GPT-5, and Gemini 3 Pro.

100 tokens per second. Two variants. One goal: agents too cheap to meter.

Are we finally at the cost floor?

[Dipak kumar Tomar @Dipaktomar73]

Sources 🧵
1. Jumper→Anthropic: https://t.co/GTD4LCmxMm
2. Huawei chips: https://t.co/psqv2v0BSl
3. Anthropic freeze: https://t.co/UVs7s0xsQI
4. Cerebras K2.6: https://t.co/Tyt1hR9OFw

Follow @Dipaktomar73 for daily AI updates. Bookmark this thread 🔖

[Dipak kumar Tomar @Dipaktomar73]

John Jumper, the Nobel Prize-winning scientist behind AlphaFold, is leaving Google DeepMind for Anthropic.

After nearly 9 years building the protein-folding system that changed biology, he joins a startup that didn't exist when he started.

+ 3 more big AI stories today 🧵

[Dipak kumar Tomar @Dipaktomar73]

Cerebras is running Kimi K2.6 at 981 tokens per second — a trillion-parameter model.

That's 6.7x faster than the next-best GPU service and 23x the median.

A 10K-token coding task completes in 5.6 seconds instead of 164.

[Dipak kumar Tomar @Dipaktomar73]

OpenAI found a way to teach AI good behavior that transfers across every domain.

A tiny dose of “beneficial trait” training improved 44 of 53 safety benchmarks.

Health-only training even made models better at catching deception.

https://t.co/cfuRpa0KfM

[Dipak kumar Tomar @Dipaktomar73]

Source: https://t.co/pFTNRFMSst…

For daily AI updates -> follow @Dipaktomar73

[Dipak kumar Tomar @Dipaktomar73]

A 3-billion-parameter model just matched a 671-billion-parameter system on olympiad math.

Weibo's VibeThinker-3B scored 94.3 on AIME 2026 — tied with DeepSeek V3.2.

It fits on a laptop. Predecessor cost $7,800 to train.

Parameter race over, or benchmarks broken?

[Dipak kumar Tomar @Dipaktomar73]

Jio just embedded an AI agent into every phone call — 500M users.

Say 'Hey Jio' during a call and it transcribes, summarizes, and even books a cab.

No app needed. 22 languages. Coming later this year.

https://t.co/wP3hmpiD2d

[Dipak kumar Tomar @Dipaktomar73]

Sources 🧵
1. AutoJack: https://t.co/dLMH95ow9E
2. Meta/Crusoe: https://t.co/m0IndNYaVo
3. Anthropic: https://t.co/0fxYDhkeGK
4. SearchLeak: https://t.co/YjE0Imsp8c

Follow @Dipaktomar73 for daily AI updates. Bookmark this thread 🔖

[Dipak kumar Tomar @Dipaktomar73]

Microsoft found a way for a malicious webpage to hijack an AI agent and execute code on your machine.

No user interaction. No credential theft. Just get the agent to open the page.

The attack is called 'AutoJack.'

+ 3 more big AI stories today 🧵

[Dipak kumar Tomar @Dipaktomar73]

Researchers found a new way to steal data from M365 Copilot — with a single link.

Dubbed 'SearchLeak.' A prompt hidden in the URL tells Copilot to read the user's email, find a 2FA code, and leak it back.

No credentials. No phishing. Just one click.