Offshore
Photo
DAIR.AI
Great paper on improving efficieny of reasoning models.
Long chain-of-thought reasoning is powerful but fundamentally limited.
The longer a model reasons, the more expensive it gets. It's well know that self-attention scales quadratically with sequence length, context windows impose hard ceilings, and critical early information fades as traces grow longer.
But what if a model could reason indefinitely without hitting any of those walls?
This new research introduces InftyThink+, an RL framework that teaches models to break reasoning into iterative rounds connected by self-generated summaries. Instead of one massive chain-of-thought, the model reasons in bounded segments, compresses its progress into a summary, and continues fresh.
Iterative reasoning only works when the model makes good decisions about when to summarize, what to preserve, and how to continue. Previous methods used supervised learning or fixed heuristics to handle these decisions. InftyThink+ treats them as a sequential decision problem optimized end-to-end with trajectory-level RL.
Training proceeds in two stages. A supervised cold-start teaches the basic iterative format. Then RL optimizes the full trajectory, learning strategic summarization and continuation policies through reward signals.
The results on DeepSeek-R1-Distill-Qwen-1.5B: InftyThink+ improves accuracy on AIME24 by 21 percentage points, outperforming conventional long chain-of-thought RL by an additional 9 points. On the out-of-distribution GPQA benchmark, it gains 5 points over the baseline and 4 points over vanilla RL. On AIME25, inference latency drops by 32.8% compared to standard reasoning. RL training itself speeds up by 18.2%.
A key finding: RL doesn't just make the model reason longer. It teaches the model to generate better summaries. When researchers replaced RL-trained summaries with external ones from a separate LLM, performance dropped. After RL training, the model's own summaries become tightly coupled with its downstream reasoning in ways external summarizers can't replicate.
The approach also decouples reasoning depth from wall-clock time. After RL, InftyThink+ extends reasoning depth while keeping latency nearly flat on several benchmarks. Standard reasoning sees latency balloon as depth increases.
Reasoning models today are bounded by context windows and crushed by quadratic attention costs. InftyThink+ removes both constraints by teaching models to reason in compressed iterations, enabling theoretically infinite-horizon reasoning with bounded compute per step.
Paper: https://t.co/VWM71BzXUf
Learn to build effective AI Agents in our academy: https://t.co/LRnpZN7L4c
tweet
Great paper on improving efficieny of reasoning models.
Long chain-of-thought reasoning is powerful but fundamentally limited.
The longer a model reasons, the more expensive it gets. It's well know that self-attention scales quadratically with sequence length, context windows impose hard ceilings, and critical early information fades as traces grow longer.
But what if a model could reason indefinitely without hitting any of those walls?
This new research introduces InftyThink+, an RL framework that teaches models to break reasoning into iterative rounds connected by self-generated summaries. Instead of one massive chain-of-thought, the model reasons in bounded segments, compresses its progress into a summary, and continues fresh.
Iterative reasoning only works when the model makes good decisions about when to summarize, what to preserve, and how to continue. Previous methods used supervised learning or fixed heuristics to handle these decisions. InftyThink+ treats them as a sequential decision problem optimized end-to-end with trajectory-level RL.
Training proceeds in two stages. A supervised cold-start teaches the basic iterative format. Then RL optimizes the full trajectory, learning strategic summarization and continuation policies through reward signals.
The results on DeepSeek-R1-Distill-Qwen-1.5B: InftyThink+ improves accuracy on AIME24 by 21 percentage points, outperforming conventional long chain-of-thought RL by an additional 9 points. On the out-of-distribution GPQA benchmark, it gains 5 points over the baseline and 4 points over vanilla RL. On AIME25, inference latency drops by 32.8% compared to standard reasoning. RL training itself speeds up by 18.2%.
A key finding: RL doesn't just make the model reason longer. It teaches the model to generate better summaries. When researchers replaced RL-trained summaries with external ones from a separate LLM, performance dropped. After RL training, the model's own summaries become tightly coupled with its downstream reasoning in ways external summarizers can't replicate.
The approach also decouples reasoning depth from wall-clock time. After RL, InftyThink+ extends reasoning depth while keeping latency nearly flat on several benchmarks. Standard reasoning sees latency balloon as depth increases.
Reasoning models today are bounded by context windows and crushed by quadratic attention costs. InftyThink+ removes both constraints by teaching models to reason in compressed iterations, enabling theoretically infinite-horizon reasoning with bounded compute per step.
Paper: https://t.co/VWM71BzXUf
Learn to build effective AI Agents in our academy: https://t.co/LRnpZN7L4c
tweet
Offshore
Video
Moon Dev
openclaw for tradingview is one of the biggest unlocks ive ever seen for traders
tradingview (and trading) will never be the same https://t.co/JVtCrJQAgG
tweet
openclaw for tradingview is one of the biggest unlocks ive ever seen for traders
tradingview (and trading) will never be the same https://t.co/JVtCrJQAgG
tweet
The Transcript
$GM CEO on their onshoring plans:
"As we look further ahead, our annual production in the U.S. is expected to rise to an industry-leading 2 million units after we begin production of the Chevrolet Equinox in Kansas, bring the Chevrolet Blazer to Tennessee and add incremental capacity for the Cadillac Escalade and launch our next-generation full-size pickups at Orion Assembly in Michigan.#
tweet
$GM CEO on their onshoring plans:
"As we look further ahead, our annual production in the U.S. is expected to rise to an industry-leading 2 million units after we begin production of the Chevrolet Equinox in Kansas, bring the Chevrolet Blazer to Tennessee and add incremental capacity for the Cadillac Escalade and launch our next-generation full-size pickups at Orion Assembly in Michigan.#
tweet
Offshore
Photo
Brady Long
RT @thisguyknowsai: 🚨 I watched a senior engineer at Anthropic build a feature in 4 hours that would've taken me 3 days.
He wasn't coding faster. He was running 8 Claude instances in parallel—each solving different parts simultaneously.
The future of coding isn't writing code. It's orchestrating AI swarms.
Here's the framework:
tweet
RT @thisguyknowsai: 🚨 I watched a senior engineer at Anthropic build a feature in 4 hours that would've taken me 3 days.
He wasn't coding faster. He was running 8 Claude instances in parallel—each solving different parts simultaneously.
The future of coding isn't writing code. It's orchestrating AI swarms.
Here's the framework:
tweet
Offshore
Photo
God of Prompt
RT @alex_prompter: R.I.P McKinsey.
You don’t need a $1,200/hr consultant anymore.
You can now run full competitive market analysis using Claude.
Here are the 10 prompts I use instead of hiring consultants: https://t.co/0G4KzDtLda
tweet
RT @alex_prompter: R.I.P McKinsey.
You don’t need a $1,200/hr consultant anymore.
You can now run full competitive market analysis using Claude.
Here are the 10 prompts I use instead of hiring consultants: https://t.co/0G4KzDtLda
tweet
Offshore
Photo
Pristine Capital
RT @realpristinecap: Hyperscalers Fueling an AI Boom!
• US Price Cycle Update 📈
• Memory Stock Trends 🧠
• Pam Bondi Testimony Wed 2/11 ⚖️
Check out tonight's research note!
https://t.co/LUxcc6ZlgM
tweet
RT @realpristinecap: Hyperscalers Fueling an AI Boom!
• US Price Cycle Update 📈
• Memory Stock Trends 🧠
• Pam Bondi Testimony Wed 2/11 ⚖️
Check out tonight's research note!
https://t.co/LUxcc6ZlgM
tweet
Offshore
Photo
Benjamin Hernandez😎
Fear and greed run the retail crowd.
I provide objective data and technical levels so you don't blow your account during volatility. Get the levels on WhatsApp.
Trade with Logic ✅ https://t.co/71FIJIdBXe
Keep your emotions off the chart.
$SOFI $HOOD $PLTR $GME $SNDK
tweet
Fear and greed run the retail crowd.
I provide objective data and technical levels so you don't blow your account during volatility. Get the levels on WhatsApp.
Trade with Logic ✅ https://t.co/71FIJIdBXe
Keep your emotions off the chart.
$SOFI $HOOD $PLTR $GME $SNDK
⚡ The "Electronic Giant" Choice
Recommendation: $AXTI ~$28.20
AXT Inc. is a "Buy" rated powerhouse with a $1.56B valuation. Today's +17.19% rally is backed by a massive 6.97M shares traded.
Reason calling it: High institutional turnover at $28.20 suggests a long-term bottom. https://t.co/dGsp8x98EG tweet
Offshore
Photo
DAIR.AI
RT @omarsar0: I think one of the most important questions in multi-agent AI right now is one almost nobody is asking: when you add more agents, are you actually getting collaboration, or are you just spending more compute?
Collaboration and communication are huge bottlenecks for multi-agent systems today.
New paper proposes a metric (Γ) that forces a distinction. You compare MAS performance against what a single agent could do with the same total resource budget. If Γ > 1, you have genuine collaboration gain. If Γ ≤ 1, you've built an expensive illusion.
Much of what gets reported as multi-agent success may just be resource accumulation. More agents means more tokens which translates to just more attempts at the problem. This is not solving for efficiency. But the bigger problem is that current benchmarks can't tell you whether the agents are actually collaborating or just brute-forcing with a bigger budget.
They also identify something AI devs will recognize: a "communication explosion" problem where unstructured agent dialogue creates so much noise that it actually suppresses collaboration below single-agent performance. More agents talking more doesn't mean more intelligence. In most cases it leads to less intelligence overall in the multi-agent system.
The metric itself is still largely aspirational. But the framing feels right. We're building multi-agent systems the way early software was built: try things, see what works, move on. The field needs something closer to a controlled experiment. Whether Γ is exactly the right lens or not, the question it forces you to ask is pointing in the right direction.
Paper: https://t.co/PKaeuZy4H5
Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX
tweet
RT @omarsar0: I think one of the most important questions in multi-agent AI right now is one almost nobody is asking: when you add more agents, are you actually getting collaboration, or are you just spending more compute?
Collaboration and communication are huge bottlenecks for multi-agent systems today.
New paper proposes a metric (Γ) that forces a distinction. You compare MAS performance against what a single agent could do with the same total resource budget. If Γ > 1, you have genuine collaboration gain. If Γ ≤ 1, you've built an expensive illusion.
Much of what gets reported as multi-agent success may just be resource accumulation. More agents means more tokens which translates to just more attempts at the problem. This is not solving for efficiency. But the bigger problem is that current benchmarks can't tell you whether the agents are actually collaborating or just brute-forcing with a bigger budget.
They also identify something AI devs will recognize: a "communication explosion" problem where unstructured agent dialogue creates so much noise that it actually suppresses collaboration below single-agent performance. More agents talking more doesn't mean more intelligence. In most cases it leads to less intelligence overall in the multi-agent system.
The metric itself is still largely aspirational. But the framing feels right. We're building multi-agent systems the way early software was built: try things, see what works, move on. The field needs something closer to a controlled experiment. Whether Γ is exactly the right lens or not, the question it forces you to ask is pointing in the right direction.
Paper: https://t.co/PKaeuZy4H5
Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX
tweet
Moon Dev
Print USDC While You Sleep: The Step-by-Step Guide to Polymarket API Bots
most people are busy gambling on sportsbooks but the real wealth is being built in the prediction markets where code handles the odds while you sleep. there is a specific bridge between being a retail trader and becoming a market maker that most people never cross because they think they need a finance degree
i believe that code is the great equalizer because i used to be the guy getting liquidated and over trading until i decided to learn to code. i spent hundreds of thousands of dollars on developers for apps in the past because i thought i was not smart enough to code myself but now fully automated systems trade for me instead of getting me wiped out
prediction markets like polymarket are not like a casino where you bet against the house but an exchange where you trade against other users. this means the prices you see are actually probabilities where an eighteen cent share represents an eighteen percent chance of an event happening
if you think the odds are better than the market suggests you buy the shares and wait for the outcome. the real secret to profiting in this space is not picking the winner of the next election but being the one who provides the liquidity for everyone else to trade
the secret of the liquidity provider is that they earn money from the spread plus the rewards the platform pays out. most traders think they have to have a crystal ball but we just need a python script that can sit on the bid and the ask simultaneously
there is a hidden danger in how these exchanges report data that can lead to your bot placing the wrong orders. i spent hours digging into the documentation to find the logic that keeps everything in sync so you do not have to repeat my mistakes
when we talk about the central limit order book we are looking at a unified system where a yes buy is essentially a no sell. once you understand that yes plus no always equals one dollar you realize that the market is always in a state of balance that can be exploited
there is a blockchain backdoor for balances that allows us to bypass the exchange api entirely when things get laggy. i discovered that we can use the web3 library in python to talk directly to the polygon rpc and see our usdc balance in real time
by looking directly at the smart contracts we can ensure that our bot always knows exactly how much capital it has to work with. this prevents the common error of trying to place an order with funds that are already locked in another trade
there is a specific logic buried in the code of top traders that allows them to make nearly a thousand dollars a day during peak volatility. i found an open source repository from a trader who was pulling in massive profits during the election and the logic is surprisingly simple
the core of this strategy is the spread scanner which looks for markets with wide gaps between the buyers and the sellers. the bot identifies the top volume markets and ranks them by the potential profit from capturing that spread while avoiding dead markets
we use a specific filter to ignore any market that is over ninety eight percent resolved because the volatility is gone and the risk is too high. the bot only focuses on the active markets where people are still fighting over the probabilities of the future
everyone talks about arbitrage but there is a hidden reason why most scanners fail and once you see it you will stop wasting time on the wrong markets. arbitrage happens when the cost of buying both yes and no is less than a dollar which is a guaranteed win
the reason most arbitrage scanners fail is that they do not account for the depth of the order book. you might see a guaranteed profit for ten cents worth of shares but as soon as you try to scale it the spread
disappears and you are left with a losing position
instead of chasing ghosts we focus on the spread bot logic which places passive bids at the best price. this allows us to ea[...]
Print USDC While You Sleep: The Step-by-Step Guide to Polymarket API Bots
most people are busy gambling on sportsbooks but the real wealth is being built in the prediction markets where code handles the odds while you sleep. there is a specific bridge between being a retail trader and becoming a market maker that most people never cross because they think they need a finance degree
i believe that code is the great equalizer because i used to be the guy getting liquidated and over trading until i decided to learn to code. i spent hundreds of thousands of dollars on developers for apps in the past because i thought i was not smart enough to code myself but now fully automated systems trade for me instead of getting me wiped out
prediction markets like polymarket are not like a casino where you bet against the house but an exchange where you trade against other users. this means the prices you see are actually probabilities where an eighteen cent share represents an eighteen percent chance of an event happening
if you think the odds are better than the market suggests you buy the shares and wait for the outcome. the real secret to profiting in this space is not picking the winner of the next election but being the one who provides the liquidity for everyone else to trade
the secret of the liquidity provider is that they earn money from the spread plus the rewards the platform pays out. most traders think they have to have a crystal ball but we just need a python script that can sit on the bid and the ask simultaneously
there is a hidden danger in how these exchanges report data that can lead to your bot placing the wrong orders. i spent hours digging into the documentation to find the logic that keeps everything in sync so you do not have to repeat my mistakes
when we talk about the central limit order book we are looking at a unified system where a yes buy is essentially a no sell. once you understand that yes plus no always equals one dollar you realize that the market is always in a state of balance that can be exploited
there is a blockchain backdoor for balances that allows us to bypass the exchange api entirely when things get laggy. i discovered that we can use the web3 library in python to talk directly to the polygon rpc and see our usdc balance in real time
by looking directly at the smart contracts we can ensure that our bot always knows exactly how much capital it has to work with. this prevents the common error of trying to place an order with funds that are already locked in another trade
there is a specific logic buried in the code of top traders that allows them to make nearly a thousand dollars a day during peak volatility. i found an open source repository from a trader who was pulling in massive profits during the election and the logic is surprisingly simple
the core of this strategy is the spread scanner which looks for markets with wide gaps between the buyers and the sellers. the bot identifies the top volume markets and ranks them by the potential profit from capturing that spread while avoiding dead markets
we use a specific filter to ignore any market that is over ninety eight percent resolved because the volatility is gone and the risk is too high. the bot only focuses on the active markets where people are still fighting over the probabilities of the future
everyone talks about arbitrage but there is a hidden reason why most scanners fail and once you see it you will stop wasting time on the wrong markets. arbitrage happens when the cost of buying both yes and no is less than a dollar which is a guaranteed win
the reason most arbitrage scanners fail is that they do not account for the depth of the order book. you might see a guaranteed profit for ten cents worth of shares but as soon as you try to scale it the spread
disappears and you are left with a losing position
instead of chasing ghosts we focus on the spread bot logic which places passive bids at the best price. this allows us to ea[...]
Offshore
Moon Dev Print USDC While You Sleep: The Step-by-Step Guide to Polymarket API Bots most people are busy gambling on sportsbooks but the real wealth is being built in the prediction markets where code handles the odds while you sleep. there is a specific bridge…
rn the liquidity rewards which are paid out in reward points for every second our orders stay live near the mid price
automation is the only way to remove the emotional baggage that comes with watching a position go against you. with bots you must iterate to success and that is why i decided to learn live so everyone can see the process of building these systems
the liquidation trap is something i know all too well from my days of manual trading and getting emotional at two in the morning. now the code handles the risk management by checking our positions every second and ensuring we never over leverage our account
we build nice functions to handle the heavy lifting like getting the token ids and placing limit orders with a single line of code. this modular approach means we can swap out strategies in seconds without having to rewrite the entire bot from scratch
the equalizer is the fact that anyone with a laptop and the willingness to learn can now compete with the big dogs. you do not need to be a math genius to understand how to bridge the gap between world events and your trading account
we are moving into a world where prediction markets will be the most accurate source of information for everything from elections to interest rates. being early to the automation side of this industry is like finding a gold mine before the rest of the world knows it exists
i am going to keep showing everything because i know that wall street will never reveal how these systems actually work. all i ask is that you stay hungry and keep iterating on your own bots until you find that edge that works for you
the final step in the process is to set up a loop that runs twenty four seven scanning for the best spreads and providing liquidity. once the system is live it becomes a machine that converts market volatility into steady growth while removing the human error that leads to liquidations
code is truly the great equalizer and once you have your first bot running you will never want to look at a chart manually ever again. the future of finance is automated and we are just getting started on this journey together
tweet
automation is the only way to remove the emotional baggage that comes with watching a position go against you. with bots you must iterate to success and that is why i decided to learn live so everyone can see the process of building these systems
the liquidation trap is something i know all too well from my days of manual trading and getting emotional at two in the morning. now the code handles the risk management by checking our positions every second and ensuring we never over leverage our account
we build nice functions to handle the heavy lifting like getting the token ids and placing limit orders with a single line of code. this modular approach means we can swap out strategies in seconds without having to rewrite the entire bot from scratch
the equalizer is the fact that anyone with a laptop and the willingness to learn can now compete with the big dogs. you do not need to be a math genius to understand how to bridge the gap between world events and your trading account
we are moving into a world where prediction markets will be the most accurate source of information for everything from elections to interest rates. being early to the automation side of this industry is like finding a gold mine before the rest of the world knows it exists
i am going to keep showing everything because i know that wall street will never reveal how these systems actually work. all i ask is that you stay hungry and keep iterating on your own bots until you find that edge that works for you
the final step in the process is to set up a loop that runs twenty four seven scanning for the best spreads and providing liquidity. once the system is live it becomes a machine that converts market volatility into steady growth while removing the human error that leads to liquidations
code is truly the great equalizer and once you have your first bot running you will never want to look at a chart manually ever again. the future of finance is automated and we are just getting started on this journey together
tweet
X (formerly Twitter)
Moon Dev (@MoonDevOnYT) on X
Print USDC While You Sleep: The Step-by-Step Guide to Polymarket API Bots
most people are busy gambling on sportsbooks but the real wealth is being built in the prediction markets where code handles the odds while you sleep. there is a specific bridge between…
most people are busy gambling on sportsbooks but the real wealth is being built in the prediction markets where code handles the odds while you sleep. there is a specific bridge between…
Michael Fritzell (Asian Century Stocks)
RT @TheYieldPig: The difference between successful people and really successful people is that really successful people say no to almost everything. - Warren Buffett
tweet
RT @TheYieldPig: The difference between successful people and really successful people is that really successful people say no to almost everything. - Warren Buffett
tweet
Moon Dev
hey you missed the zoom call
you missed our private zoom call today
it is an absolute must see if you are trying to build trading bots
get a ticket for tomorrows zoom and you will get the replay from today
here is the link https://t.co/JbJdIbW2p9
see you tomorrow
moon dev
tweet
hey you missed the zoom call
you missed our private zoom call today
it is an absolute must see if you are trying to build trading bots
get a ticket for tomorrows zoom and you will get the replay from today
here is the link https://t.co/JbJdIbW2p9
see you tomorrow
moon dev
tweet