Byron Tang

learn everyday, applied AI eng

[Byron Tang @tangbyron]

Continual learning for LLMs and agents will be critical for the next step-function improvement.

Ran experiments on recursive language models + code execution, and found ~6× gains over ReAct on BrowseComp Plus, a long context, multi-hop benchmark.

https://t.co/xK0i6RUTtg

[Byron Tang @tangbyron]

Love the investment and post from @PrimeIntellect , and great work from @a1zhang @lateinteraction. I ran my own experiments with RLM on BrowseComp-Plus, and saw similar huge gains in accuracy.

https://t.co/CaYzawabNK

The original blog post in Oct (a lifetime ago in AI, of course) from Alex is a great read. The simplicity "felt" right, and maps well to the bitter lesson.

[Omar Khattab @lateinteraction]

Excellent bet as usual from the @PrimeIntellect folks.

Stay tuned for lots more from @a1zhang in the very near future… :-)

[Byron Tang @tangbyron]

@peakji congrats Manus team! epic launch, $100m arr, and now joining Meta. cheering for yall.

[Byron Tang @tangbyron]

Love it! Simple and elegant solution. question though - cc best practices say to clear context often, so when you mentioned cc runs for days, does the stop hook plugin ever clear context, or just lean on compaction? Or is there a hybrid where the stop hook calls a subagent to document progress in a md, clear context, and trigger cc to continue based on progress md

[Byron Tang @tangbyron]

Christmas came a bit earlier this year! Great job by the @pollenrobotics and @huggingface team for putting together a great product in Reachy Mini

A few initial thoughts:
- assembly does take 2-3 hrs, but is superbly documented with a video companion. Had a lot of fun in the process
- perhaps as a function of building it, the hardware feels sturdy and will last. Solid rods and motors for the 6 degrees of freedom movement. Plenty of expressiveness!
- can get started immediately through the spaces app store, the mini conversation app is great out of the box. Super low latency, my kids already love it. Also comes natively multimodal

Can’t wait to try out the python sdk next!

[Byron Tang @tangbyron]

@simpsoka agreed. 2026 we will see agentic actions triggered from chat -> proactively / opportunistically delivered. and as a result 10x the number of actions and impact.

[Byron Tang @tangbyron]

Leverage both client and server side context filtering to aggressively protect the context for the primary agent.

We can go beyond conversational level compaction on the client side, and leverage "Sub-LM", or isolated LLM sessions to summarize large tool results. This offers more precise compression.

[Byron Tang @tangbyron]

Sharing a few tips on context engineering for tool heavy agents with Claude Agent SDK https://t.co/Wzon0fYZNz

Specifically, designing the system for three levels of disclosure
- Agents need high reasoning density
- Users want transparency and visuals
- Logs require structured output to accelerate development, especially when leveraging code execution to chain tool calls

[Byron Tang @tangbyron]

@AnthropicAI I was inspired by the great RLM post from @a1zhang , and Reasoning Bank paper from @GoogleAI @Siru_Ouyang
- rlm https://t.co/ERE58ucAw2
- reasoning bank https://t.co/nzwG1uuqqw

[Byron Tang @tangbyron]

Continual learning for LLMs and agents will be critical for the next step-function improvement.

Ran experiments on recursive language models + code execution, and found ~6× gains over ReAct on BrowseComp Plus, a long context, multi-hop benchmark.

https://t.co/xK0i6RUTtg

[Byron Tang @tangbyron]

This echos the @AnthropicAI post on reducing context rot through code execution https://t.co/WTzOpOFxUM

The additional twist is adding the Sub-LM to compress and reflect in deep research tasks. In ablations, skipping this step caused large context overflows and thrashing

[Byron Tang @tangbyron]

This echos the @AnthropicAI post on reducing context rot through code execution https://t.co/WTzOpOFxUM

The additional twist is adding the Sub-LM to compress and reflect in deep research tasks. In ablations, skipping this step caused large context overflows and thrashing

[Byron Tang @tangbyron]

It was so interesting to see the Root-LM come up with retrieval, derivation, and filtering strategies that would have taken multiple turns in the traditional Agent <> Tool interactions. Essentially using Python as a planning language rather than step-by-step tool calls.