🤖 AI - Transformer Weights and Post-Training Changes

🤖 AI - Transformer Weights and Post-Training Changes

Post-training modifications to transformer weights can alter the model's behavior, but what specifically changes to make it no longer a language model? Pretraining, self-supervised fine-tuning, preference optimization, and reinforcement learning all modify the same weights used by the inference process that gives us a conditional distribution.

Key Points:

• Transformer Weights and Post-Training Changes: Post-training modifications can alter the model's behavior, but the underlying weights remain the same. The inference process that gives us a conditional distribution is unchanged.

• Trade-offs and Failure Modes: The trade-offs and failure modes of post-training modifications are not well understood, and more research is needed to fully grasp their implications.

• Actionable Takeaway: When modifying transformer weights post-training, it's essential to understand the underlying mechanisms and potential trade-offs to avoid unintended consequences.

🔗 Resources:

• Original post ↗
• Original source
• AlexTensor
• Discussion on transformer weights and post-training changes

Image

🚨 Urgent: Help Needed for Stranded Family Member

My brother Vijayababu Shanmugam passed away in Hanoi on Sep24. My in-law is stranded alone at Hanoi in shock. I request urgent embassy support and help to bring him home. @MEAIndia @IndiainVietnam @CMOTamilnadu @DrSJaishankar

Key Points:

• Urgent Situation: A family member has passed away, and the in-law is stranded alone in Hanoi.

• Request for Help: Urgent embassy support and help are needed to bring the in-law home.

• Actionable Takeaway: If you or someone you know is in a similar situation, reach out to the relevant authorities and seek assistance.

🔗 Resources:

• Original post ↗
• Original source
• Sureshkannascbe
• Request for help and support

Image

🚨 ThinkPad Build Quality Decline

This is largely true but the build quality of ThinkPads has declined considerably in the past 5-7 years. I have multiple T series that are 8+ years old with no problems. My last three T & X units have failed with regular use (screens, MBs, etc).

Key Points:

• Decline in Build Quality: The build quality of ThinkPads has declined significantly in the past 5-7 years.

• Comparison to Older Models: Older ThinkPad models (8+ years old) have no problems, while newer models (T & X series) have failed with regular use.

• Actionable Takeaway: When purchasing a ThinkPad, consider the age and build quality of the model.

🔗 Resources:

• Original post ↗
• Original source
• kscottz
• Discussion on ThinkPad build quality decline

Image

🚨 T Series ThinkPads and Regular Use

I used to toss my T old series in a backpack or panier and bike across town without issue. The new ones seem to fail with just regular use and largely sitting on a desk.

Key Points:

• Comparison to Older Models: Older ThinkPad models (T series) were more durable and could withstand regular use, while newer models fail with minimal use.

• Actionable Takeaway: When purchasing a ThinkPad, consider the age and build quality of the model.

🔗 Resources:

• Original post ↗
• Original source
• kscottz
• Discussion on T series ThinkPads and regular use

Image

🚨 Pessimism in Basic Research

One of the worst forms of brainrot AI has cultivated is pessimism about basic research, ie the idea that important work can only happen inside a frontier/neo lab and only with 10k+ GPUs, so the rest should not even bother. What a bleak way to think about science. And it's false.

Key Points:

• Pessimism in Basic Research: AI has cultivated pessimism about basic research, leading to a bleak view of science.

• Counter-Argument: Basic research can be done outside of frontier/neo labs and with fewer resources.

• Actionable Takeaway: Encourage and support basic research, regardless of the resources available.

🔗 Resources:

• Original post ↗
• Original source
• DimitrisPapail
• Discussion on pessimism in basic research

Image

🚨 Drinking and Driving

Maybe it’s a Ukrainian speaking in me (zero alcohol tolerance when driving, by law), but it’s such a turn-off when a guy wants to come to a drinking date (“let’s share a bottle of wine, shall we?”) in a car :/ Like, how are you getting home? Driving your tipsy ass? Spare me.

Key Points:

• Drinking and Driving: Drinking and driving is a serious issue, and it's essential to prioritize safety.

• Actionable Takeaway: If you plan to drink, have a designated driver or alternative transportation.

🔗 Resources:

• Original post ↗
• Original source
• mtmthh
• Discussion on drinking and driving

Image

🚀 IROS2026 Workshop on Perception and Decision Making

At #IROS2026 today for our workshop on Perception and Decision Making for Athletic Humanoid Robotics! We’ll have a panel discussion on open problems in athletic humanoid robotics at 4 PM. Come join us if you’re at IROS! https:// iros-2026-athletic-humanoid.github.io/workshop/

Key Points:

• IROS2026 Workshop: The workshop will cover Perception and Decision Making for Athletic Humanoid Robotics.

• Panel Discussion: A panel discussion on open problems in athletic humanoid robotics will take place at 4 PM.

• Actionable Takeaway: Attend the workshop and panel discussion to learn more about athletic humanoid robotics.

🔗 Resources:

• Original post ↗
• Original source
• ZiyunClaudeWang
• IROS2026 workshop and panel discussion

Image

🚨 John Paul II Statue and Reviewer 3

Fun fact: when this John Paul II statue was unveiled in 2011, Romans hated it so much it got sent back for revisions and re-unveiled the next year. Even a pope gets Reviewer 3 !

Key Points:

• John Paul II Statue: The John Paul II statue was unveiled in 2011 but was later sent back for revisions due to public dislike.

• Reviewer 3: Even the Pope is subject to review and criticism.

• Actionable Takeaway: No one is immune to criticism, not even the Pope.

🔗 Resources:

• Original post ↗
• Original source
• silvirouskin
• Discussion on John Paul II statue and Reviewer 3

Image

🚨 Inference and Model Scaling

When it comes to inference, the dedicated chapter for it in the book "How to scale your model" can't be missed. https:// jax-ml.github.io/scaling-book/i nference/ …

Key Points:

• Inference and Model Scaling: Inference is a critical component of model scaling, and the dedicated chapter in the book provides valuable insights.

• Actionable Takeaway: Read the dedicated chapter on inference in the book "How to scale your model" to learn more about model scaling.

🔗 Resources:

• Original post ↗
• Original source
• vivekgalatage
• Inference and model scaling

Image

Image

🚨 GPT-6 Astra and Real-Time Robotics

People watching GPT-6 Astra demos on X may not realise how slow it actually is in real time. Here’s a rollout of GPT-6 Astra controlling the MolmoAct2 YAMs to complete this task in real-time. There is potential, but i don’t think it’s practical for real-world robotics deployment

Key Points:

• GPT-6 Astra and Real-Time Robotics: GPT-6 Astra is slow in real-time, but it has potential for robotics deployment.

• Actionable Takeaway: GPT-6 Astra may not be practical for real-world robotics deployment due to its slow performance.

🔗 Resources:

• Original post ↗
• Original source
• DJiafei
• GPT-6 Astra and real-time robotics

Image