Offshore
Photo
Brady Long
RT @thisguyknowsai: I reverse-engineered the actual prompting frameworks that top AI labs use internally.
Not the fluff you see on Twitter.
The real shit that turns vague inputs into precise, structured outputs.
Spent 3 weeks reading OpenAI's model cards, Anthropic's constitutional AI papers, and leaked internal prompt libraries.
Here's what actually moves the needle:
tweet
RT @thisguyknowsai: I reverse-engineered the actual prompting frameworks that top AI labs use internally.
Not the fluff you see on Twitter.
The real shit that turns vague inputs into precise, structured outputs.
Spent 3 weeks reading OpenAI's model cards, Anthropic's constitutional AI papers, and leaked internal prompt libraries.
Here's what actually moves the needle:
tweet
Offshore
Photo
The Transcript
$ABNB CEO: Airbnb’s defense against disintermediation is focusing on what AI can’t replicate
"A chatbot can give you a list of homes, but it can't give you the unique ones you find on Airbnb..." https://t.co/5lwQ6BVXcD
tweet
$ABNB CEO: Airbnb’s defense against disintermediation is focusing on what AI can’t replicate
"A chatbot can give you a list of homes, but it can't give you the unique ones you find on Airbnb..." https://t.co/5lwQ6BVXcD
tweet
The Transcript
RT @TheTranscript_: Microsoft commits to building frontier in-house foundationreducing OpenAI dependence
"We have to develop our own foundation models, which are at the absolute frontier, with gigawatt-scale compute and some of the very best AI training teams in the world" - $MSFT AI chief
[FT]
tweet
RT @TheTranscript_: Microsoft commits to building frontier in-house foundationreducing OpenAI dependence
"We have to develop our own foundation models, which are at the absolute frontier, with gigawatt-scale compute and some of the very best AI training teams in the world" - $MSFT AI chief
[FT]
tweet
Offshore
Photo
DAIR.AI
// Improving Efficiency of Evolutionary AI Agents //
Evolutionary AI agents are powerful but can be wasteful.
Systems, inspired by AlphaEvolve and OpenEvolve, iteratively generate, mutate, and refine candidate solutions using LLMs. However, every refinement step invokes the same large model regardless of task difficulty.
Most mutations don't need a 32B model.
This new research introduces AdaptEvolve, a framework that dynamically selects which model handles each evolutionary step based on intrinsic generation confidence.
Instead of routing everything through the largest available model, a lightweight decision tree router estimates whether the small model's output is sufficient or needs escalation.
The confidence signal comes from four entropy-based metrics computed on the small model's token probabilities: Mean Confidence for global assurance, Lowest Group Confidence for localized reasoning collapses, Tail Confidence for solution stability, and Bottom-K% Confidence for distinguishing noise from systematic hallucination.
A shallow decision tree, bootstrapped from just 50 warm-up examples, uses these signals to make real-time routing decisions.
What makes this practical?
The router adapts online. An Adaptive Hoeffding Tree continuously updates its decision boundaries as the evolutionary population drifts toward harder edge cases.
On LiveCodeBench, AdaptEvolve retains 97.9% of the 32B upper-bound accuracy (73.6% vs 75.2%) while cutting compute cost by 34.4%. On MBPP, the router identifies that 85% of queries are solvable by the 4B model alone, reducing cost by 41.5% while maintaining 97.1% of peak accuracy. Across benchmarks, the method reduces total inference compute by 37.9% while retaining 97.5% of the upper-bound performance.
Evolutionary agents don't need maximum capability at every step. Confidence-driven routing turns the cost-capability trade-off from a fixed choice into a dynamic, per-step decision.
Paper: https://t.co/YSNCKZuTeN
Learn to build effective AI Agents in our academy: https://t.co/LRnpZN7L4c
tweet
// Improving Efficiency of Evolutionary AI Agents //
Evolutionary AI agents are powerful but can be wasteful.
Systems, inspired by AlphaEvolve and OpenEvolve, iteratively generate, mutate, and refine candidate solutions using LLMs. However, every refinement step invokes the same large model regardless of task difficulty.
Most mutations don't need a 32B model.
This new research introduces AdaptEvolve, a framework that dynamically selects which model handles each evolutionary step based on intrinsic generation confidence.
Instead of routing everything through the largest available model, a lightweight decision tree router estimates whether the small model's output is sufficient or needs escalation.
The confidence signal comes from four entropy-based metrics computed on the small model's token probabilities: Mean Confidence for global assurance, Lowest Group Confidence for localized reasoning collapses, Tail Confidence for solution stability, and Bottom-K% Confidence for distinguishing noise from systematic hallucination.
A shallow decision tree, bootstrapped from just 50 warm-up examples, uses these signals to make real-time routing decisions.
What makes this practical?
The router adapts online. An Adaptive Hoeffding Tree continuously updates its decision boundaries as the evolutionary population drifts toward harder edge cases.
On LiveCodeBench, AdaptEvolve retains 97.9% of the 32B upper-bound accuracy (73.6% vs 75.2%) while cutting compute cost by 34.4%. On MBPP, the router identifies that 85% of queries are solvable by the 4B model alone, reducing cost by 41.5% while maintaining 97.1% of peak accuracy. Across benchmarks, the method reduces total inference compute by 37.9% while retaining 97.5% of the upper-bound performance.
Evolutionary agents don't need maximum capability at every step. Confidence-driven routing turns the cost-capability trade-off from a fixed choice into a dynamic, per-step decision.
Paper: https://t.co/YSNCKZuTeN
Learn to build effective AI Agents in our academy: https://t.co/LRnpZN7L4c
tweet
Offshore
Photo
God of Prompt
RT @godofprompt: CLAUDE IS SO COOKED THIS TIME
China just dropped Kimi K2.5, the best open model for OpenClaw(ClawdBot)
It's on par with Claude Opus4.5,
but 8x CHEAPER!!!
It's currently the #1 most used model for OpenClaw and the #1 most used model overall on OpenRouter!
Here's everything you should know:
tweet
RT @godofprompt: CLAUDE IS SO COOKED THIS TIME
China just dropped Kimi K2.5, the best open model for OpenClaw(ClawdBot)
It's on par with Claude Opus4.5,
but 8x CHEAPER!!!
It's currently the #1 most used model for OpenClaw and the #1 most used model overall on OpenRouter!
Here's everything you should know:
tweet
Offshore
Video
DAIR.AI
RT @omarsar0: Just incredible that this is possible today.
One of my favorite MCP tools as of late.
Just prompt to generate beautiful excalidraw diagrams. https://t.co/YgH57NOAoT
tweet
RT @omarsar0: Just incredible that this is possible today.
One of my favorite MCP tools as of late.
Just prompt to generate beautiful excalidraw diagrams. https://t.co/YgH57NOAoT
tweet
Offshore
Photo
Fiscal.ai
Amazon's e-commerce business is more profitable than it has ever been.
9% Operating Margins in North America.
How much more profitable can this segment get?
$AMZN https://t.co/ZOmpWA5z55
tweet
Amazon's e-commerce business is more profitable than it has ever been.
9% Operating Margins in North America.
How much more profitable can this segment get?
$AMZN https://t.co/ZOmpWA5z55
tweet
Offshore
Video
Illiquid
Ben offers a great intro to Taiwan. I also agree that Substacks should be bundled or rolled up, and tried to do it myself.
Interestingly, at close to the 6 month mark, I’m only slightly behind Ben when he first started, despite having a much smaller target audience. I have two weeks to catch him!
tweet
Ben offers a great intro to Taiwan. I also agree that Substacks should be bundled or rolled up, and tried to do it myself.
Interestingly, at close to the 6 month mark, I’m only slightly behind Ben when he first started, despite having a much smaller target audience. I have two weeks to catch him!
The tech world closely reads @benthompson's Stratechery, and for good reason. On Cheeky Pint, we discuss ads coming to AI, the Saaspocalypse, TSMC, and the media business. (Plus: how to visit Taiwan.)
00:00:20 Visiting Taiwan
00:04:59 Aggregation and AI
00:23:53 TikTok/Bytedance
00:29:58 Aggregation and AI redux
00:35:31 Agentic commerce
00:45:08 Is SaaS canceled?
00:52:21 Stratechery
01:03:36 How Ben uses AI
01:06:06 The TSMC break
01:13:53 Rapid fire
01:20:53 Feedback on Stripe tweet
Offshore
Photo
Pristine Capital
RT @realpristinecap: A massive regime change is underway in 2026🔄
For the first time in years, the "Growth at any price" trade is taking a backseat. We are seeing a powerful rotation into Value and Small Caps as market breadth expands 👇 https://t.co/W51cleqVR1
tweet
RT @realpristinecap: A massive regime change is underway in 2026🔄
For the first time in years, the "Growth at any price" trade is taking a backseat. We are seeing a powerful rotation into Value and Small Caps as market breadth expands 👇 https://t.co/W51cleqVR1
tweet
Offshore
Photo
App Economy Insights
📊 This Week in Visuals
$CSCO $SHOP $AMAT $AZN $KO $MCD $TMUS $ANET $APP $NTES $HOOD $ABNB $MAR $HLT $HERMES $RACE $F $NET $DDOG $COIN $ADYEY $FISV $KHC $EXPE $QSR $TWLO $TOST $Z $DKNG $PINS $HUBS $LYFT $KVYO $MNDY
https://t.co/ZAyrJNoouN
tweet
📊 This Week in Visuals
$CSCO $SHOP $AMAT $AZN $KO $MCD $TMUS $ANET $APP $NTES $HOOD $ABNB $MAR $HLT $HERMES $RACE $F $NET $DDOG $COIN $ADYEY $FISV $KHC $EXPE $QSR $TWLO $TOST $Z $DKNG $PINS $HUBS $LYFT $KVYO $MNDY
https://t.co/ZAyrJNoouN
tweet
Offshore
Photo
App Economy Insights
🗓️ What are you watching this week?
• Tuesday: $PANW $CDNS
• Wednesday: $ADI $MCO $FIG $BKNG $GPN
• Thursday: $MELI $NU $WMT $BABA $KLAR $LYV
All visualized in our newsletter! https://t.co/aALlk4Yk2n
tweet
🗓️ What are you watching this week?
• Tuesday: $PANW $CDNS
• Wednesday: $ADI $MCO $FIG $BKNG $GPN
• Thursday: $MELI $NU $WMT $BABA $KLAR $LYV
All visualized in our newsletter! https://t.co/aALlk4Yk2n
tweet