The best way to learn about LLMs is by building your own from scratch and I think this repo is the best for it.
Did anyone try it? Let's connect if u r also interested in ai and tech! 🤝
https://t.co/ietJCuMXET
I had an interview yesterday for SDE-1, got reject in screen round only. I though everything went great but received the mail today that they are moving with other candidates.
I know rejections are part of the journey but still it's hard man.
#rejection#sde#job
Bro worked at multiple top companies, sold his startup to mercor in just 9 months. Has 280k+ follower on LinkedIn, 210k+ on youtube, and is working on multiple opensource projects on the side.
He have achieved what I am aiming for.
Looking to connect with people interested in:
Software Engineering
Backend, frontend
full-stack
System Design, LLD
I am learning these and will be posting what I learn.
Let's connect! ☺️
It's been 3 days since my last post.
And full detox mode is still on.
Since then I applied to 30+ companies, revised JavaScript, Typescript, some fundamentals and projects.
Next up practice DSA and machine coding.
#detox#JobSearch#growth#motivation
I have already tried it for at least 1.5 days. And sometimes I just get these urges to watch something or listen to music. My will is the only thing keeping me in control. So I realised, I might need something more to control these cravings and this post is the result of that.
From today I am going full detox mode for 1 month.
No Music
No Movies or Shows
No reels
No porn or masterbating.
I will be posting the effects, results, challenges here itself.
#SelfControl#selfimprovement#detox
GPT-6 Astra is live.
Not “better chatbot.” A model that can sit on a computer and finish the work.
OpenAI just launched GPT-6 Astra — SOTA on computer use, coding, science, and professional workflows.
The numbers:
• ARC-AGI-3: 99.9%
• FrontierMath Tier 4: 98%
• ExploitBench: 100%
• 1.05M context
• API: GPT-6 Astra at $10 / $50 per 1M tokens
Greg Brockman’s line: welcome to the AGI era.
Rollout is messy (limited orgs first, then Plus / Pro / Business / Enterprise, API, Azure, Bedrock). That’s the story. The model is the bigger story.
If you build software, this is the week to test agents on real tasks — PRs, research, forms, multi-step desktop work — not another toy prompt.
Launch: https://t.co/lo4gZn0S5S
Safety: https://t.co/NSYAuJjT06
API:https://t.co/kOov3gH6tP
Claude Fable 5.1 shipped yesterday.
What’s actually new: same list price as Fable 5, cache reads cut 75% to $0.25 per million tokens, and Anthropic is pitching it for long coding and knowledge-work sessions. Early users are already talking about cache windows, usage limits, and whether the cheaper reads change how they run Claude Code.
How LLMs are actually built (explained in 60 seconds)
Blank algorithm → tokenization → self-attention → pre-training on next-token prediction → fine-tuning…
and it starts talking like you.
Clean visual breakdown, zero jargon:
This video was made with AI in single prompt.
Want to know exactly how? Comment “HOW” and I’ll DM you.
Quick extras this week:
Taiwan indicted 9 people over Nvidia AI server smuggling to China
Australia banned fully AI-generated songs from the official music charts
🚨 AI just had a wild 3-4 days. Here’s the rapid-fire roundup:
Nvidia eyeing $30B+ stake in Perplexity (ARR already $750M)
Groq 3 LPX hits production — 3,400 tokens/sec for agents
Alibaba Wan3.0 drops: 30-sec videos from text or PDFs
Thomson Reuters launches its own frontier legal LLM for just $40M
Hardware, capital, and domain models all racing. The future just got closer.
Read the thread for more details. 🧵👇
5. OpenAI’s GPT-5.6 family (Sol / Terra / Luna) is now live inside AWS’s Kiro
Spec-driven agentic coding environment. Joint testing showed major cost reductions on complex tasks.
Source: OpenAI + Kiro
https://t.co/IlOm2fqzYo
https://t.co/bgQHNdMuy1
4. Thomson Reuters launched its own proprietary frontier LLM called “Thomson”
Trained for ~$40M on decades of Westlaw, Practical Law & Reuters content.
Competitive with top frontier models on legal tasks, fully owned and controlled by the company. First use: CoCounsel Legal.
Source: Official Thomson Reuters press release
https://t.co/MdHuYGUeoX
3. Alibaba officially launched Wan3.0
Native 30-second video generation (double the previous limit) from text, images, and documents/PDFs/slides/webpages.
“Everything to video” is here.
Source: Reuters + Alibaba Cloud
https://t.co/ftFl00q1TR
https://t.co/3l8RHcyFuV
2. Nvidia’s Groq 3 LPX inference accelerator is now in full production
Built from the $20B Groq deal, it hits a record 3,400 tokens/sec on long-context agentic workloads (Gemma 4 31B, 100k context).
Up to 256 accelerators per rack. Nebius is the first customer.
Also detailed: the 88-core Vera CPU at Hot Chips.
Source: NVIDIA + SiliconANGLE
https://t.co/GbZP68w2sb
https://t.co/F3mgRLha4m
1. Nvidia is in talks to invest in Perplexity at a $30B+ valuation
That’s more than 50% higher than its last round.
Perplexity’s annualized revenue has jumped past $750M (from under $250M at the start of 2026), driven by its agentic tool “Perplexity Computer.”
Source: Reuters / The Information
https://t.co/LJqiaWDxWt