Tahoe Therapeutics has built AI bio's largest virtual cell model. And it's giving it away for free.
Tahoe-x1 is a foundation AI model trained to simulate entire cell behaviors, including gene expression, drug responses, and interactions across biological contexts. It’s touted as the largest of its kind in AI-driven biotechnology and is open-source to foster community research.
The model was trained on over 300 million unique cells, incorporating:
- Tahoe-100M а dataset of 100 million cells with 60,000 drug-cell interactions (1,200 molecules tested on 50 cancer cell lines).
This was a "world record" at the time, built with Parse Biosciences (sample prep) and Ultima Genomics (sequencing).
- Future Goals: Tahoe aims to scale to 1 billion single-cell datapoints and 1 million drug-patient interactions, likened to a "GPT moment" for biology, akin to large language models for text.
Models built on Tahoe-100M (e.g., by Arc Institute) showed 2x better accuracy than competitors. Tahoe-x1 advances this further by integrating diverse data across species, tissues, and perturbations (external stimuli like drugs).
Key Achievements and Partnerships
1. Tahoe-100M contributed to the Arc Virtual Cell Atlas (300+ million cells), a public resource launched with Arc Institute in February 2025, making data accessible to researchers worldwide.
2. Tahoe-x1 has identified new drug candidates for cancer (including "undruggable" targets) and novel therapeutic targets. The company is advancing its pipeline toward clinical trials and seeking one strategic partner (pharma or AI company) for co-development.
3. Tahoe competes with initiatives like the Chan Zuckerberg Initiative, which also aims for 1 billion cell datasets for AI modeling.
Tahoe-x1 is a foundation AI model trained to simulate entire cell behaviors, including gene expression, drug responses, and interactions across biological contexts. It’s touted as the largest of its kind in AI-driven biotechnology and is open-source to foster community research.
The model was trained on over 300 million unique cells, incorporating:
- Tahoe-100M а dataset of 100 million cells with 60,000 drug-cell interactions (1,200 molecules tested on 50 cancer cell lines).
This was a "world record" at the time, built with Parse Biosciences (sample prep) and Ultima Genomics (sequencing).
- Future Goals: Tahoe aims to scale to 1 billion single-cell datapoints and 1 million drug-patient interactions, likened to a "GPT moment" for biology, akin to large language models for text.
Models built on Tahoe-100M (e.g., by Arc Institute) showed 2x better accuracy than competitors. Tahoe-x1 advances this further by integrating diverse data across species, tissues, and perturbations (external stimuli like drugs).
Key Achievements and Partnerships
1. Tahoe-100M contributed to the Arc Virtual Cell Atlas (300+ million cells), a public resource launched with Arc Institute in February 2025, making data accessible to researchers worldwide.
2. Tahoe-x1 has identified new drug candidates for cancer (including "undruggable" targets) and novel therapeutic targets. The company is advancing its pipeline toward clinical trials and seeking one strategic partner (pharma or AI company) for co-development.
3. Tahoe competes with initiatives like the Chan Zuckerberg Initiative, which also aims for 1 billion cell datasets for AI modeling.
Endpoints News
Exclusive: In the race to build virtual cells, AI bio Tahoe open-sources its own model
Tahoe Therapeutics is unveiling Tahoe-x1, the largest virtual cell AI model, and will release it as open-source. CEO Nima Alidoust leads the 15-employee biotech's initiative.
👏7🔥3🆒2🥰1
Ant group introduced Ring-1T, a 1T-parameter MoE reasoning model with ~50B params active per token.
It’s trained with a long-CoT SFT phase, a verifiable-rewards reasoning RL phase, then a general RLHF phase, and introduces three pieces that make trillion-scale RL actually run:
- IcePop to stabilize updates
- C3PO++ to keep GPUs busy under a token budget
- ASystem to unify high-throughput RL stack
On benchmarks, it leads open weights on AIME-25, HMMT-25, ARC-AGI-1, LiveCodeBench, CodeForces, and ArenaHard v2.
It reaches silver-medal level on IMO-2025 using only natural-language reasoning.
It’s trained with a long-CoT SFT phase, a verifiable-rewards reasoning RL phase, then a general RLHF phase, and introduces three pieces that make trillion-scale RL actually run:
- IcePop to stabilize updates
- C3PO++ to keep GPUs busy under a token budget
- ASystem to unify high-throughput RL stack
On benchmarks, it leads open weights on AIME-25, HMMT-25, ARC-AGI-1, LiveCodeBench, CodeForces, and ArenaHard v2.
It reaches silver-medal level on IMO-2025 using only natural-language reasoning.
🔥5👍2👏2🆒1
Microsoft introduced 12 new Copilot features. With Groups, Copilot goes multiplayer.
You can now collaborate in real time with your team + Copilot to brainstorm, co-write, plan, or study together.
Plan and collaborate with the new Groups feature—perfect for organizing study sessions, family vacations, or a night out. Copilot makes your crew more productive by answering questions, assigning tasks, and getting everyone moving.
More engaging conversations with Copilot through the new Mico appearance. It brings that much more personal experience to the notion of your AI companion.
And, shop smarter online with the new Edge, AI browser. Copilot, with your permission, can evaluate your open tabs to make more confident decisions and then take action to help you book reservations, travel, or make smarter shopping decisions.
You can now collaborate in real time with your team + Copilot to brainstorm, co-write, plan, or study together.
Plan and collaborate with the new Groups feature—perfect for organizing study sessions, family vacations, or a night out. Copilot makes your crew more productive by answering questions, assigning tasks, and getting everyone moving.
More engaging conversations with Copilot through the new Mico appearance. It brings that much more personal experience to the notion of your AI companion.
And, shop smarter online with the new Edge, AI browser. Copilot, with your permission, can evaluate your open tabs to make more confident decisions and then take action to help you book reservations, travel, or make smarter shopping decisions.
Microsoft Copilot Blog
Human-centered AI | Microsoft Copilot Blog
Today, we’re dropping the Copilot Fall Release, a big step forward in making AI more personal, useful, and human-centered. There’s a lot of noise around AI. Headlines, hype, fear. At Microsoft AI, we want to change the outlook. We’re betting on optimism…
👍2🔥2👏2🥴2
You spend $1B training a model A.
Someone on your team leaves and launches their own model API B.
You're suspicious. Was B was derived (e.g., fine-tuned) from A?
But you only have blackbox access to B.
With this paper, you can still tell with strong statistical guarantees (p-values < 1e-8).
Idea: test for independence of A's training data order with likelihoods under B.
There are crazy amounts of metadata about training process baked into the model that can't be washed out, like a palimpsest.
Someone on your team leaves and launches their own model API B.
You're suspicious. Was B was derived (e.g., fine-tuned) from A?
But you only have blackbox access to B.
With this paper, you can still tell with strong statistical guarantees (p-values < 1e-8).
Idea: test for independence of A's training data order with likelihoods under B.
There are crazy amounts of metadata about training process baked into the model that can't be washed out, like a palimpsest.
👍3🔥2🥰2
Stanford and Tsinghua University presented Ctrl-World is a controllable world model that generalizes zero-shot to new environments, cameras, and objects.
Build on pre-trained video model, Ctrl-World adds key designs to make it compatible with modern VLA:
1) Fully controllable via low-level action conditioning.
2) Multi-view prediction including wrist-view.
3) Context as memory for consistency.
Despite these promising progress, also notice WM can still:
- Fail in modeling complex physical interactions.
- Sensitive to initial observations.
Build on pre-trained video model, Ctrl-World adds key designs to make it compatible with modern VLA:
1) Fully controllable via low-level action conditioning.
2) Multi-view prediction including wrist-view.
3) Context as memory for consistency.
Despite these promising progress, also notice WM can still:
- Fail in modeling complex physical interactions.
- Sensitive to initial observations.
ctrl-world.github.io
Ctrl-World
👍3🥰2👏2
All about AI, Web 3.0, BCI
Tahoe Therapeutics has built AI bio's largest virtual cell model. And it's giving it away for free. Tahoe-x1 is a foundation AI model trained to simulate entire cell behaviors, including gene expression, drug responses, and interactions across biological…
Tahoe introduced Tahoe-x1 (Tx1) a 3B parameter, single-cell foundation model that learns unified representations of genes, cells, and drugs, achieving state-of-the-art performance across cancer-relevant cell biology benchmarks, open-sourced on HuggingFace.
Tx1 is the first billion-parameter, compute-efficient foundation model trained on perturbation-rich single-cell data. And it is fully open-source with open weights.
Tx1 is a Transformer (scGPT-inspired self-supervised objective) that stays practical to train at 3B parameter scale, enabling empirical search for optimal architectures and hyperparameters for modeling cells. It is 3-30x more compute efficient than other cell state models.
To make it so, researchers borrowed the best tricks from LMs (FlashAttention v2, FSDP, Dataset streaming, Mixed-precision + multi-node scaling). But even cooler: researchers improved the attention operation at the heart of these models.
Researchers designed new benchmarks to evaluate its performance in cancer-relevant discovery and translational tasks.
Tx1 achieves SOTA on key discovery tasks. It outperforms other models (and for the first time matches with linear baselines) in predicting gene essentiality as measured by the landmark DepMap dataset, a key piece of data in identifying subtype-specific targets.
Tx1 is the first billion-parameter, compute-efficient foundation model trained on perturbation-rich single-cell data. And it is fully open-source with open weights.
Tx1 is a Transformer (scGPT-inspired self-supervised objective) that stays practical to train at 3B parameter scale, enabling empirical search for optimal architectures and hyperparameters for modeling cells. It is 3-30x more compute efficient than other cell state models.
To make it so, researchers borrowed the best tricks from LMs (FlashAttention v2, FSDP, Dataset streaming, Mixed-precision + multi-node scaling). But even cooler: researchers improved the attention operation at the heart of these models.
Researchers designed new benchmarks to evaluate its performance in cancer-relevant discovery and translational tasks.
Tx1 achieves SOTA on key discovery tasks. It outperforms other models (and for the first time matches with linear baselines) in predicting gene essentiality as measured by the landmark DepMap dataset, a key piece of data in identifying subtype-specific targets.
huggingface.co
tahoebio/Tahoe-x1 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
👍4❤3🔥2
New work Anthropic and Thinking Machines
AI companies use model specifications to define desirable behaviors during training. Are model specs clearly expressing what we want models to do? And do different frontier models have different personalities?
The problem: specifications can be inherently ambiguous, or scenarios can force tradeoffs between competing principles, causing models to make wildly different choices.
The experiments show that high disagreement among frontier models strongly correlates with specification issues—indicating important gaps in current behavioral guidelines.
Datasets.
AI companies use model specifications to define desirable behaviors during training. Are model specs clearly expressing what we want models to do? And do different frontier models have different personalities?
The problem: specifications can be inherently ambiguous, or scenarios can force tradeoffs between competing principles, causing models to make wildly different choices.
The experiments show that high disagreement among frontier models strongly correlates with specification issues—indicating important gaps in current behavioral guidelines.
Datasets.
huggingface.co
jifanz/stress_testing_model_spec · Datasets at Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
❤4🔥3👏2
MiniMax’s M2 achieves a new all-time-high Intelligence Index score for an open weights model and offers impressive efficiency with only 10B active parameters (200B total)
Key takeaways:
1. Efficiency to serve at scale: MiniMax-M2 has 200B total parameters and is very sparse with only 10B active parameters per forward pass. Such few active parameters allow the model to be served efficiently at scale (DeepSeek V3.2 has 671B total and 37B active, Qwen3 has 235B total and 22B active). The model can also easily fit on 4xH100s at FP8 precision
2. Strengths focus on agentic use-cases: The model’s strengths include tool use and instruction following (as shown by Tau2 Bench and IFBench). As such, while M2 likely excels at agentic use cases it may underperform other open weights leaders such as DeepSeek V3.2 and Qwen3 235B at some generalist tasks. This is in line with a number of recent open weights model releases from Chinese AI labs which focus on agentic capabilities, likely pointing to a heavy post-training emphasis on RL. Similar to most other leading open weights models, M2 is a text only model - Alibaba’s recent Qwen3 VL releases remain the leading open weights multimodal models
3. Cost & token usage: MiniMax’s API is offering the model at a very competitive per token price of $0.3/$1.2 per 1M input/output tokens. However, the model is very verbose, using 120M token to complete our Intelligence Index evaluations - equal highest along with Grok 4. As such, while it is a low priced model this is moderated by high token usage
4. Continued leadership in open source by Chinese AI labs: MiniMax’s release continues the leadership of Chinese AI labs in open source that DeepSeek kicked off in late 2024, and which has been continued by continued DeepSeek releases, Alibaba, Z AI and Moonshot AI
Key takeaways:
1. Efficiency to serve at scale: MiniMax-M2 has 200B total parameters and is very sparse with only 10B active parameters per forward pass. Such few active parameters allow the model to be served efficiently at scale (DeepSeek V3.2 has 671B total and 37B active, Qwen3 has 235B total and 22B active). The model can also easily fit on 4xH100s at FP8 precision
2. Strengths focus on agentic use-cases: The model’s strengths include tool use and instruction following (as shown by Tau2 Bench and IFBench). As such, while M2 likely excels at agentic use cases it may underperform other open weights leaders such as DeepSeek V3.2 and Qwen3 235B at some generalist tasks. This is in line with a number of recent open weights model releases from Chinese AI labs which focus on agentic capabilities, likely pointing to a heavy post-training emphasis on RL. Similar to most other leading open weights models, M2 is a text only model - Alibaba’s recent Qwen3 VL releases remain the leading open weights multimodal models
3. Cost & token usage: MiniMax’s API is offering the model at a very competitive per token price of $0.3/$1.2 per 1M input/output tokens. However, the model is very verbose, using 120M token to complete our Intelligence Index evaluations - equal highest along with Grok 4. As such, while it is a low priced model this is moderated by high token usage
4. Continued leadership in open source by Chinese AI labs: MiniMax’s release continues the leadership of Chinese AI labs in open source that DeepSeek kicked off in late 2024, and which has been continued by continued DeepSeek releases, Alibaba, Z AI and Moonshot AI
🔥6❤5🥰4
Google introduced next-generation conversational agents, including:
1 - a new, easy-to-use low-code visual builder
2 - natural-sounding voices
3 - reduced deployment time
4 - unified governance
1 - a new, easy-to-use low-code visual builder
2 - natural-sounding voices
3 - reduced deployment time
4 - unified governance
Google Cloud Blog
Introducing Gemini Enterprise | Google Cloud Blog
Today, we’re introducing Gemini Enterprise – the new front door for AI in the workplace. It’s our advanced agentic platform that brings the best of Google AI to every employee, for every workflow.
🔥5👏5❤2
Neuralink announced the first participant in the UK
Paul, who is paralyzed due to motor neuron disease, received his Neuralink implant at UCLH earlier this month and was able to control a computer with his thoughts just hours after surgery.
Paul, who is paralyzed due to motor neuron disease, received his Neuralink implant at UCLH earlier this month and was able to control a computer with his thoughts just hours after surgery.
NHS
First UK patient uses thought to control computer hours after Neuralink implant
A patient with motor neurone disease was able to control a computer just by using his thoughts following the UK’s first Neuralink implant surgery at UCLH’s National Hospital for Neurology and Neurosurgery (NHNN) in October 2025.
👏5❤3🔥2
Clifford_Chance_x_Deutsche_Bank_blockchainrf.pdf
19.5 MB
Clifford Chance and Deutsche Bank looked at the growing intersection of AI and DLT in a new report.
AI predictability and DLT immutability could create autonomous systems that move beyond efficiency to self-verifiability, capable of executing, auditing and learning simultaneously.
AI-augmented smart contracts translate business logic into executable code through LLMs, lowering technical barriers and accelerating deployment. Deutsche Bank’s own pilot with finaXai for tokenized fund servicing illustrates how AI can automate code audits and stress-test logic across multi-jurisdictional workflows.
AI-powered blockchain oracles enhance market-data integrity and anomaly detection, evolving from passive data relays to predictive control layers. The Chainlink–Euroclear–Swift initiative demonstrates how AI-driven oracle networks could synchronise fragmented corporate-action data, cutting validation costs that currently exceed USD 3–5 million annually across custodians.
Tokenized data marketplaces use DLT to enable verifiable AI training through traceable data provenance and federated learning allowing hospitals or banks to train models locally without exposing raw data. This architecture could unlock high-value private datasets while aligning with GDPR, the EU AI Act, and MiCA’s data-governance provisions.
Agentic AI with blockchain wallets gives autonomous systems controlled financial agency. By combining multi-signature wallets with programmable limits, these agents could conduct low-value transactions independently, laying the groundwork for AI-to-AI commerce within a USD 36 trillion digital-payments economy, from automated insurance claims to real-time treasury optimization
Yet convergence amplifies regulatory tension.
AI predictability and DLT immutability could create autonomous systems that move beyond efficiency to self-verifiability, capable of executing, auditing and learning simultaneously.
AI-augmented smart contracts translate business logic into executable code through LLMs, lowering technical barriers and accelerating deployment. Deutsche Bank’s own pilot with finaXai for tokenized fund servicing illustrates how AI can automate code audits and stress-test logic across multi-jurisdictional workflows.
AI-powered blockchain oracles enhance market-data integrity and anomaly detection, evolving from passive data relays to predictive control layers. The Chainlink–Euroclear–Swift initiative demonstrates how AI-driven oracle networks could synchronise fragmented corporate-action data, cutting validation costs that currently exceed USD 3–5 million annually across custodians.
Tokenized data marketplaces use DLT to enable verifiable AI training through traceable data provenance and federated learning allowing hospitals or banks to train models locally without exposing raw data. This architecture could unlock high-value private datasets while aligning with GDPR, the EU AI Act, and MiCA’s data-governance provisions.
Agentic AI with blockchain wallets gives autonomous systems controlled financial agency. By combining multi-signature wallets with programmable limits, these agents could conduct low-value transactions independently, laying the groundwork for AI-to-AI commerce within a USD 36 trillion digital-payments economy, from automated insurance claims to real-time treasury optimization
Yet convergence amplifies regulatory tension.
👍5🔥2👏2
Huge breakthrough from Google DeepMind. The study was led by AlphaGo’s creator, David Silver.
In their latest Nature paper, “Discovering SOTA reinforcement learning algorithms,” they show that AI can autonomously discover better RL algorithms.
"Enabling machines to discover learning algorithms for themselves is one of the most promising ideas in AI."
Could the next generation of RL algorithms be machine-discovered?
In their latest Nature paper, “Discovering SOTA reinforcement learning algorithms,” they show that AI can autonomously discover better RL algorithms.
"Enabling machines to discover learning algorithms for themselves is one of the most promising ideas in AI."
Could the next generation of RL algorithms be machine-discovered?
Nature
Discovering state-of-the-art reinforcement learning algorithms
Nature - An autonomous method discovers reinforcement learning rules from the cumulative experiences of a population of agents across a large number of complex environments, and the discovered rule...
🔥7🥰2👏2
Google rolled out a Gemini-powered personal health coach inside Fitbit. It uses a deep-agent architecture to orchestrate between conversational, data science, and domain expert sub-agents.
- Performs complex numerical reasoning on physiological time series data.
- Available to eligible U.S. Android Fitbit Premium users, with iOS expanding soon.
- Validated via 1 million+ human annotations and 100k+ hours of evaluation.
- Personalized guidance via a 5-10 minute interactive text or voice conversation.
- Adaptive plans based on individual health metrics and goals.
- Grounded in behavioral science and Consumer Health Advisory Panel.
- Performs complex numerical reasoning on physiological time series data.
- Available to eligible U.S. Android Fitbit Premium users, with iOS expanding soon.
- Validated via 1 million+ human annotations and 100k+ hours of evaluation.
- Personalized guidance via a 5-10 minute interactive text or voice conversation.
- Adaptive plans based on individual health metrics and goals.
- Grounded in behavioral science and Consumer Health Advisory Panel.
research.google
How we are building the personal health coach
👍6🔥2🥰2
Western Union to build stablecoin on Solana blockchain, issue with Anchorage
Western Union, the global remittance giant processing around 70 million cross-border transactions quarterly across 200+ countries, has announced plans to launch its own dollar-backed stablecoin in the first half of 2026.
The token, dubbed USDPT (U.S. Dollar Payment Token), will be built on the Solana blockchain. Solana was chosen for its high-speed, low-cost transactions—ideal for remittances—potentially enabling near-instant settlements and slashing fees compared to legacy systems.
Expected in H1 2026, starting as a pilot in select corridors (e.g., South America and Africa) before broader rollout.
Use Cases:
- Faster cross-border transfers with real-time settlement.
- Improved FX pricing in low-liquidity markets.
- Customer custody options, acting like a "savings account" in USD for users in high-inflation regions.
Western Union is also developing a digital wallet network with third-party providers to let users buy, sell, and hold stablecoins directly. CEO Devin McGranahan emphasized stablecoins as an "opportunity, not a threat," positioning the company to compete with rivals like PayPal (PYUSD) and MoneyGram.
Western Union, the global remittance giant processing around 70 million cross-border transactions quarterly across 200+ countries, has announced plans to launch its own dollar-backed stablecoin in the first half of 2026.
The token, dubbed USDPT (U.S. Dollar Payment Token), will be built on the Solana blockchain. Solana was chosen for its high-speed, low-cost transactions—ideal for remittances—potentially enabling near-instant settlements and slashing fees compared to legacy systems.
Expected in H1 2026, starting as a pilot in select corridors (e.g., South America and Africa) before broader rollout.
Use Cases:
- Faster cross-border transfers with real-time settlement.
- Improved FX pricing in low-liquidity markets.
- Customer custody options, acting like a "savings account" in USD for users in high-inflation regions.
Western Union is also developing a digital wallet network with third-party providers to let users buy, sell, and hold stablecoins directly. CEO Devin McGranahan emphasized stablecoins as an "opportunity, not a threat," positioning the company to compete with rivals like PayPal (PYUSD) and MoneyGram.
The Wall Street Journal
Exclusive | Western Union, Early Telegraph Pioneer, Joins the Crypto Arms Race
The company plans to launch its own stablecoin in 2026 to send money around the globe, a move that might lower customer costs and settle transactions faster.
🔥5👏4❤2⚡1
OpenAI is planning to build an AI Cloud Platform. "More value created by people building on the platform than by the platform builder".
OpenAI have a clear line of sight to automating AI research. Chief Scientist Jakub Pachocki said that OpenAI expects to be able to develop an automated AI research intern by September 2026 and fully automated AI researcher by March 2028.
OpenAI have a clear line of sight to automating AI research. Chief Scientist Jakub Pachocki said that OpenAI expects to be able to develop an automated AI research intern by September 2026 and fully automated AI researcher by March 2028.
Openai
Livestream
🔥3👍2🥰2
Ant group presents the 1st trillion-scale open-source thinking model
Key highlights:
- Superior thinking performance 93.40 on AIME25, 86.72 on HMMT25, 2088 on CodeForces, 55.94 on ARC-AGI-1, plus a silver medal on IMO 2025
- IcePop Mitigates the training-inference mismatch issue in MoE RL, ensuring stable and growing RL training
- C3PO++ Introduces a dynamic rollout partitioning mechanism, achieving 2.5x faster inference and 1.5x faster training
- ASystem a high-performance RL framework designed for efficiently scaling training to trillion-parameter thinking models
Key highlights:
- Superior thinking performance 93.40 on AIME25, 86.72 on HMMT25, 2088 on CodeForces, 55.94 on ARC-AGI-1, plus a silver medal on IMO 2025
- IcePop Mitigates the training-inference mismatch issue in MoE RL, ensuring stable and growing RL training
- C3PO++ Introduces a dynamic rollout partitioning mechanism, achieving 2.5x faster inference and 1.5x faster training
- ASystem a high-performance RL framework designed for efficiently scaling training to trillion-parameter thinking models
arXiv.org
Every Step Evolves: Scaling Reinforcement Learning for...
We present Ring-1T, the first open-source, state-of-the-art thinking model with a trillion-scale parameter. It features 1 trillion total parameters and activates approximately 50 billion per...
Salesforce introduced MMPersuade, a comprehensive multimodal benchmark that assesses AI agents’ susceptibility to established persuasion principles, covering commercial, subjective and behavioral, and adversarial contexts.
MMPersuade is a new dataset and evaluation framework to systematically study multimodal persuasion in LVLMs.
Team built a comprehensive multimodal benchmark pairing persuasive strategies with over 62,000 images and 4,700 videos.
It covers 3 key contexts: Commercial (Sales & Ads), Subjective & Behavioral (Health Nudging, Politics) , Adversarial (Misinformation & Fabricated Claims)
MMPersuade is a new dataset and evaluation framework to systematically study multimodal persuasion in LVLMs.
Team built a comprehensive multimodal benchmark pairing persuasive strategies with over 62,000 images and 4,700 videos.
It covers 3 key contexts: Commercial (Sales & Ads), Subjective & Behavioral (Health Nudging, Politics) , Adversarial (Misinformation & Fabricated Claims)
Carnegie, Stanford introduced a new work on Training LLMs to Discover Abstractions for Solving Reasoning Problems
cohenqu.github.io
RLAD: RL through Abstraction Discovery
🔥2🥰2👏2🆒2
MIT presented LoRA vs full fine-tuning: same performance ≠ same solution.
This paper shows that LoRA and full fine-tuning, even when equally well fit, learn structurally different solutions and that LoRA forgets less and can be made even better (lesser forgetting) by a simple intervention.
This paper shows that LoRA and full fine-tuning, even when equally well fit, learn structurally different solutions and that LoRA forgets less and can be made even better (lesser forgetting) by a simple intervention.
👍2🥰2👏2
New Anthropic research: Signs of introspection in LLMs.
Can language models recognize their own internal thoughts? Or do they just make up plausible answers when asked about them?
Anthropic found evidence for genuine—though limited—introspective capabilities in Claude.
Researchers developed a method to distinguish true introspection from made-up answers: inject known concepts into a model's “brain,” then see how these injections affect the model’s self-reported internal states.
In one experiment, researchers asked the model to detect when a concept is injected into its “thoughts.” When researchers inject a neural pattern representing a particular concept, Claude can in some cases detect the injection, and identify the concept.
However, it doesn’t always work. In fact, most of the time, models fail to exhibit awareness of injected concepts, even when they are clearly influenced by the injection.
Also show that Claude introspects in order to detect artificially prefilled outputs. Normally, Claude apologizes for such outputs. But if researchers retroactively inject a matching concept into its prior activations, team can fool Claude into thinking the output was intentional.
This reveals a mechanism that checks consistency between intention and execution. The model appears to compare "what did I plan to say?" against "what actually came out?"—a form of introspective monitoring happening in natural circumstances.
Also found evidence for cognitive control, where models deliberately "think about" something. For instance, when team instruct a model to think about "aquariums” in an unrelated context, researchers measure higher aquarium-related neural activity than if team instruct it not to.
Note that experiments do not address the question of whether AI models can have subjective experience or human-like self-awareness. The mechanisms underlying the behaviors observe are unclear, and may not have the same philosophical significance as human introspection.
While currently limited, AI models’ introspective capabilities will likely grow more sophisticated. Introspective self-reports could help improve the transparency of AI models’ decision-making—but should not be blindly trusted.
Can language models recognize their own internal thoughts? Or do they just make up plausible answers when asked about them?
Anthropic found evidence for genuine—though limited—introspective capabilities in Claude.
Researchers developed a method to distinguish true introspection from made-up answers: inject known concepts into a model's “brain,” then see how these injections affect the model’s self-reported internal states.
In one experiment, researchers asked the model to detect when a concept is injected into its “thoughts.” When researchers inject a neural pattern representing a particular concept, Claude can in some cases detect the injection, and identify the concept.
However, it doesn’t always work. In fact, most of the time, models fail to exhibit awareness of injected concepts, even when they are clearly influenced by the injection.
Also show that Claude introspects in order to detect artificially prefilled outputs. Normally, Claude apologizes for such outputs. But if researchers retroactively inject a matching concept into its prior activations, team can fool Claude into thinking the output was intentional.
This reveals a mechanism that checks consistency between intention and execution. The model appears to compare "what did I plan to say?" against "what actually came out?"—a form of introspective monitoring happening in natural circumstances.
Also found evidence for cognitive control, where models deliberately "think about" something. For instance, when team instruct a model to think about "aquariums” in an unrelated context, researchers measure higher aquarium-related neural activity than if team instruct it not to.
Note that experiments do not address the question of whether AI models can have subjective experience or human-like self-awareness. The mechanisms underlying the behaviors observe are unclear, and may not have the same philosophical significance as human introspection.
While currently limited, AI models’ introspective capabilities will likely grow more sophisticated. Introspective self-reports could help improve the transparency of AI models’ decision-making—but should not be blindly trusted.
Anthropic
Emergent introspective awareness in large language models
Research from Anthropic on the ability of large language models to introspect
❤3🔥2👏2
Cognition (ex-team of Windsurf) released SWE-1.5, fast agent model.
That delivers "near-SOTA coding performance" at significantly higher speeds.
U can try it here.
That delivers "near-SOTA coding performance" at significantly higher speeds.
U can try it here.
Cognition
Cognition | Introducing SWE-1.5: Our Fast Agent Model
Today we’re releasing SWE-1.5, the latest in our family of models optimized for software engineering. It is a frontier-size model with hundreds of billions of parameters that achieves near-SOTA coding performance. It also sets a new standard for speed: we…
❤2🔥2👏2