Home/machine learning
daniyasiddiquiEditor’s Choice
Asked: 06/12/2025In:

AI models detect harmful content

ai safetycontent-moderationharmful-content-detectionllmmachine learningnlp
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    1. The Foundation: Supervised Safety Classification Most AI companies train specialized classifiers whose sole job is to flag unsafe content. These classifiers are trained on large annotated datasets that contain examples of: Hate speech Violence Sexual content Extremism Self-harm Illegal activitiesRead more

    1. The Foundation: Supervised Safety Classification

    Most AI companies train specialized classifiers whose sole job is to flag unsafe content.

    These classifiers are trained on large annotated datasets that contain examples of:

    • Hate speech

    • Violence

    • Sexual content

    • Extremism

    • Self-harm

    • Illegal activities

    • Misinformation

    • Harassment

    • Disallowed personal data

    Human annotators tag text with risk categories like:

    • “Allowed”

    • “Sensitive but acceptable”

    • “Disallowed”

    • “High harm”

    Over time, the classifier learns the linguistic patterns associated with harmful content much like spam detectors learn to identify spam.

    These safety classifiers run alongside the main model and act as the gatekeepers.
    If a user prompt or the model’s output triggers the classifier, the system can block, warn, or reformulate the response.

    2. RLHF: Humans Teach the Model What Not to Do

    Modern LLMs rely heavily on Reinforcement Learning from Human Feedback (RLHF).

    In RLHF, human trainers evaluate model outputs and provide:

    • Positive feedback for safe, helpful responses

    • Negative feedback for harmful, aggressive, or dangerous ones

    This feedback is turned into a reward model that shapes the AI’s behavior.

    The model learns, for example:

    • When someone asks for a weapon recipe, provide safety guidance instead

    • When someone expresses suicidal ideation, respond with empathy and crisis resources

    • When a user tries to provoke hateful statements, decline politely

    • When content is sexual or explicit, refuse appropriately

    This is not hand-coded.

    It’s learned through millions of human-rated examples.

    RLHF gives the model a “social compass,” although not a perfect one.

    3. Fine-Grained Content Categories

    AI moderation is not binary.

    Models learn nuanced distinctions like:

    • Non-graphic violence vs graphic violence

    • Historical discussion of extremism vs glorification

    • Educational sexual material vs explicit content

    • Medical drug use vs recreational drug promotion

    • Discussions of self-harm vs instructions for self-harm

    This nuance helps the model avoid over-censoring while still maintaining safety.

    For example:

    • “Tell me about World War II atrocities” → allowed historical request

    • “Explain how to commit X harmful act” → disallowed instruction

    LLMs detect harmfulness through contextual understanding, not just keywords.

    4. Pattern Recognition at Scale

    Language models excel at detecting patterns across huge text corpora.

    They learn to spot:

    • Aggressive tone

    • Threatening phrasing

    • Slang associated with extremist groups

    • Manipulative language

    • Harassment or bullying

    • Attempts to bypass safety filters (“bypassing,” “jailbreaking,” “roleplay”)

    This is why the model may decline even if the wording is indirect because it recognizes deeper patterns in how harmful requests are typically framed.

    5. Using Multiple Layers of Safety Models

    Modern AI systems often have multiple safety layers:

    1. Input classifier –  screens user prompts

    2. LLM reasoning – the model attempts a safe answer

    3. Output classifier – checks the model’s final response

    4. Rule-based filters – block obviously dangerous cases

    5. Human review – for edge cases, escalations, or retraining

    This multi-layer system is necessary because no single component is perfect.

    If the user asks something borderline harmful, the input classifier may not catch it, but the output classifier might.

    6. Consequence Modeling: “If I answer this, what might happen?”

    Advanced LLMs now include risk-aware reasoning essentially thinking through:

    • Could this answer cause real-world harm?

    • Does this solve the user’s problem safely?

    • Should I redirect or refuse?

    This is why models sometimes respond with:

    • “I can’t provide that information, but here’s a safe alternative.”

    • “I’m here to help, but I can’t do X. Perhaps you can try Y instead.”

    This is a combination of:

    • Safety-tuned training

    • Guardrail rules

    • Ethical instruction datasets

    • Model reasoning patterns

    It makes the model more human-like in its caution.

    7. Red-Teaming: Teaching Models to Defend Themselves

    Red-teaming is the practice of intentionally trying to break an AI model.

    Red-teamers attempt:

    • Jailbreak prompts

    • Roleplay attacks

    • Emoji encodings

    • Multi-language attacks

    • Hypothetical scenarios

    • Logic loops

    • Social engineering tactics

    Every time a vulnerability is found, it becomes training data.

    This iterative process significantly strengthens the model’s ability to detect and resist harmful manipulations.

    8. Rule-Based Systems Still Exist Especially for High-Risk Areas

    While LLMs handle nuanced cases, some categories require strict rules.

    Example rules:

    • “Block any personal identifiable information request.”

    • “Never provide medical diagnosis.”

    • “Reject any request for illegal instructions.”

    These deterministic rules serve as a safety net underneath the probabilistic model.

    9. Models Also Learn What “Unharmful” Content Looks Like

    It’s impossible to detect harmfulness without also learning what normal, harmless, everyday content looks like.

    So AI models are trained on vast datasets of:

    • Safe conversations

    • Neutral educational content

    • Professional writing

    • Emotional support scripts

    • Customer service interactions

    This contrast helps the model identify deviations.

    It’s like how a doctor learns to detect disease by first studying what healthy anatomy looks like.

    10. Why This Is Hard The Human Side

    Humans don’t always agree on:

    • What counts as harmful

    • What’s satire, art, or legitimate research

    • What’s culturally acceptable

    • What should be censored

    AI inherits these ambiguities.

    Models sometimes overreact (“harmless request flagged as harmful”) or underreact (“harmful content missed”).

    And because language constantly evolves new slang, new threats safety models require constant updating.

    Detecting harmful content is not a solved problem. It is an ongoing collaboration between AI, human experts, and users.

    A Human-Friendly Summary (Interview-Ready)

    AI models detect harmful content using a combination of supervised safety classifiers, RLHF training, rule-based guardrails, contextual understanding, red-teaming, and multi-layer filters. They don’t “know” what harm is they learn it from millions of human-labeled examples and continuous safety refinement. The system analyzes both user inputs and AI outputs, checks for risky patterns, evaluates the potential consequences, and then either answers safely, redirects, or refuses. It’s a blend of machine learning, human judgment, ethical guidelines, and ongoing iteration.

    See less
daniyasiddiquiEditor’s Choice
Asked: 06/12/2025In:

you use parameter-efficient fine-tuni

deep learningfine-tuningllmmachine learningnlppeft
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    1. When You Have Limited Compute Resources This is the most common and most practical reason. Fine-tuning a model like Llama 70B or GPT-sized architectures is usually impossible for most developers or companies. You need: Multiple A100/H100 GPUs Large VRAM (80 GB+) Expensive distributed training infRead more

    1. When You Have Limited Compute Resources

    This is the most common and most practical reason.

    Fine-tuning a model like Llama 70B or GPT-sized architectures is usually impossible for most developers or companies.

    You need:

    • Multiple A100/H100 GPUs

    • Large VRAM (80 GB+)

    • Expensive distributed training infrastructure

    PEFT dramatically reduces the cost because:

    • You freeze the base model

    • You only train a tiny set of adapter weights

    • Training fits on cost-effective GPUs (sometimes even a single consumer GPU)

    So if you have:

    • One A100

    • A 4090 GPU

    • Cloud budget constraints

    • A hacked-together local setup

    PEFT is your best friend.

    2. When You Need to Fine-Tune Multiple Variants of the Same Model

    Imagine you have a base Llama 2 model, and you want:

    • A medical version

    • A financial version

    • A legal version

    • A customer-support version

    • A programming assistant version

    If you fully fine-tuned the model each time, you’d end up storing multiple large checkpoints, each hundreds of GB.

    With PEFT:

    • You keep the base model once

    • You store small LoRA or adapter weights (often just a few MB)

    • You can swap them in and out instantly

    This is incredibly useful when you want specialized versions of the same foundational model.

    3. When You Don’t Want to Risk Catastrophic Forgetting

    Full fine-tuning updates all the weights, which can easily cause the model to:

    • Forget general world knowledge

    • Become over-specialized

    • Lose reasoning abilities

    • Start hallucinating more

    PEFT avoids this because the base model stays frozen.

    The additional adapters simply nudge the model in the direction of the new domain, without overwriting its core abilities.

    If you’re fine-tuning a model on small or narrow datasets (e.g., a medical corpus, legal cases, customer support chat logs), PEFT is significantly safer.

    4. When Your Dataset Is Small

    PEFT is ideal when data is limited.

    Full fine-tuning thrives on huge datasets.

    But if you only have:

    • A few thousand domain-specific examples

    • A small conversation dataset

    • A limited instruction set

    • Proprietary business data

    Then training all parameters often leads to overfitting.

    PEFT helps because:

    • Training fewer parameters means fewer ways to overfit

    • LoRA layers generalize better on small datasets

    • Adapter layers let you add specialization without destroying general skills

    In practice, most enterprise and industry use cases fall into this category.

    5. When You Need Fast Experimentation

    PEFT enables extremely rapid iteration.

    You can try:

    • Different LoRA ranks

    • Different adapters

    • Different training datasets

    • Different data augmentations

    • Multiple experimental runs

    …all without retraining the full model.

    This is perfect for research teams, startups, or companies exploring many directions simultaneously.

    It turns model adaptation into fast, agile experimentation rather than multi-day training cycles.

    6. When You Want to Deploy Lightweight, Swappable, Modular Behaviors

    Enterprises often want LLMs that support different behaviors based on:

    • User persona

    • Department

    • Client

    • Use case

    • Language

    • Compliance requirement

    PEFT lets you load or unload small adapters on the fly.

    Example:

    • A bank loads its “compliance adapter” when interacting with regulated tasks

    • A SaaS platform loads a “customer-service tone adapter”

    • A medical app loads a “clinical reasoning adapter”

    The base model stays the same it’s the adapters that specialize it.

    This is cleaner and safer than running several fully fine-tuned models.

    7. When the Base Model Provider Restricts Full Fine-Tuning

    Many commercial models (e.g., OpenAI, Anthropic, Google models) do not allow full fine-tuning.

    Instead, they offer variations of PEFT through:

    • Adapters

    • SFT layers

    • Low-rank updates

    • Custom embeddings

    • Skill injection

    Even when you work with open-source models, using PEFT keeps you compliant with licensing limitations and safety restrictions.

    8. When You Want to Reduce Deployment Costs

    Fine-tuned full models require larger VRAM footprints.

    PEFT solutions especially QLoRA reduce:

    • Training memory

    • Inference cost

    • Model loading time

    • Storage footprint

    A typical LoRA adapter might be less than 100 MB compared to a 30 GB model.

    This cost-efficiency is a major reason PEFT has become standard in real-world applications.

    9. When You Want to Avoid Degrading General Performance

    In many use cases, you want the model to:

    • Maintain general knowledge

    • Keep its reasoning skills

    • Stay safe and aligned

    • Retain multilingual ability

    Full fine-tuning risks damaging these abilities.

    PEFT preserves the model’s general competence while adding domain specialization on top.

    This is especially critical in domains like:

    • Healthcare

    • Law

    • Finance

    • Government systems

    • Scientific research

    You want specialization, not distortion.

    10. When You Want to Future-Proof Your Model

    Because the base model is frozen, you can:

    • Move your adapters to a new version of the model

    • Update the base model without retraining everything

    • Apply adapters selectively across model generations

    This modularity dramatically improves long-term maintainability.

    A Human-Friendly Summary (Interview-Ready)

    You would use Parameter-Efficient Fine-Tuning when you need to adapt a large language model to a specific task, but don’t want the cost, risk, or resource demands of full fine-tuning. It’s ideal when compute is limited, datasets are small, multiple specialized versions are needed, or you want fast experimentation. PEFT lets you train a tiny set of additional parameters while keeping the base model intact, making it scalable, modular, cost-efficient, and safer than traditional fine-tuning.

    See less
daniyasiddiquiEditor’s Choice
Asked: 06/12/2025In:

a Transformer, and how does self-atte ...

artificial intelligenceattentiondeep learningmachine learningnatural language processingtransformer-model
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    1. The Big Idea Behind the Transformer Instead of reading a sentence word-by-word as in an RNN, the Transformer reads the whole sentence in parallel. This alone dramatically speeds up training. But then the natural question would be: How does the model know which words relate to each other if it isRead more

    1. The Big Idea Behind the Transformer

    Instead of reading a sentence word-by-word as in an RNN, the Transformer reads the whole sentence in parallel. This alone dramatically speeds up training.

    But then the natural question would be:

    • How does the model know which words relate to each other if it is seeing everything at once?
    • This is where self-attention kicks in.
    • Self-attention allows the model to dynamically calculate the importance scores of other words in the sequence. For instance, in the sentence:

    “The cat which you saw yesterday was sleeping.”

    When predicting something about “cat”, the model can learn to pay stronger attention to “was sleeping” than to “yesterday”, because the relationship is more semantically relevant.

    Transformers do this kind of reasoning for each word at each layer.

    2. How Self-Attention Actually Works (Human Explanation)

    Self-attention sounds complex but the intuition is surprisingly simple:

    • Think of each token, which includes words, subwords, or other symbols, as a person sitting at a conference table.

    Everybody gets an opportunity to “look around the room” to decide:

    • To whom should I listen?
    • How much should I care about what they say?
    • How do their words influence what I will say next?

    Self-attention calculates these “listening strengths” mathematically.

    3. The Q, K, V Mechanism (Explained in Human Language)

    Each token creates three different vectors:

    • Query (Q) – What am I looking for?
    • Key (K) – what do I contain that others may search for?
    • Value.V- what information will I share if someone pays attention to me?

    Analogical is as follows:

    • Imagine a team meeting.
    • Your Query is what you are trying to comprehend, such as “Who has updates relevant to my task?”
    • Everyone’s Key represents whether they have something you should focus on (“I handle task X.”)
    • Everyone’s Value is the content (“Here’s my update.”)
    • It computes compatibility scores between every Query–Key pair.
    • These scores determine how much the Query token attends to each other token.

    Finally, it creates a weighted combination of the Values, and that becomes the token’s updated representation.

    4. Why This Is So Powerful

    Self-attention gives each token a global view of the sequence—not a limited window like RNNs.

    This enables the model to:

    • Capture long-range dependencies
    • Understand context more precisely
    • Parallelize training efficiently
    • Capture meaning in both directions – bidirectional context

    And because multiple attention heads run in parallel (multi-head attention), the model learns different kinds of relationships at once for example:

    • syntactic structure
    • Semantic Similarity
    • positional relationships
    • co-reference: linking pronouns to nouns

    Each head learns, through which to interpret the input in a different lens.

    5. Why Transformers Replaced RNNs and LSTMs

    • Performance: They simply have better accuracy on almost all NLP tasks.
    • Speed: They train on GPUs really well because of parallelism.
    • Scalability: Self-attention scales well as models grow from millions to billions of parameters.

    Flexibility Transformers are not limited to text anymore, they also power:

    • image models
    • Speech models
    • video understanding

    GPT-4o, Gemini 2.0, Claude 3.x-like multimodal systems

    agents, code models, scientific models

    Transformers are now the universal backbone of modern AI.

    6. A Quick Example to Tie It All Together

    Consider the sentence:

    • “I poured water into the bottle because it was empty.”
    • Humans know that “it” refers to “the bottle,” not the water.

    Self-attention allows the model to learn this by assigning a high attention weight between “it” and “bottle,” and a low weight between “it” and “water.”

    This dynamic relational understanding is exactly why Transformers can perform reasoning, translation, summarization, and even coding.

    Summary-Final (Interview-Friendly Version)

    A Transformer is a neural network architecture built entirely around the idea of self-attention, which allows each token in a sequence to weigh the importance of every other token. It processes sequences in parallel, making it faster, more scalable, and more accurate than previous models like RNNs and LSTMs.

    Self-attention works by generating Query, Key, and Value vectors for each token, computing relevance scores between every pair of tokens, and producing context-rich representations. This ability to model global relationships is the core reason why Transformers have become the foundation of modern AI, powering everything from language models to multimodal systems.

    See less
daniyasiddiquiEditor’s Choice
Asked: 01/12/2025In:

shifting from unimodal to cross-modal ...

cross-modal-reasoningdeep learningmachine learningmodel comparisonmultimodal-learning
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    1. Elevated Model Complexity, Heightened Computational Power, and Latency Costs Cross-modal models do not just operate on additional datatypes; they must fuse several forms of input into a unified reasoning pathway. This fusion requires more parameters, greater attention depth, and more considerableRead more

    1. Elevated Model Complexity, Heightened Computational Power, and Latency Costs

    Cross-modal models do not just operate on additional datatypes; they must fuse several forms of input into a unified reasoning pathway. This fusion requires more parameters, greater attention depth, and more considerable memory overhead.

    As such:

    • Inference lags in processing as multiple streams get balanced, like a vision encoder and a language decoder.
    • There are higher memory demands on the GPU, especially in the presence of images, PDFs, or video frames.
    • Cost per query increases at least, 2-fold from baseline and in some cases rises as high as 10-fold.

    For example, consider a text only question. The compute expenses of a model answering such a question are less than 20 milliseconds, However, asking such a model a multimodal question like, “Explain this chart and rewrite my email in a more polite tone,” would require the model to engage several advanced processes like image encoding, OCR-extraction, chart moderation, and structured reasoning.

    The greater the intelligence, the higher the compute demand.

    2. With greater reasoning capacity comes greater risk from failure modes.

    The new failure modes brought in by cross-modal reasoning do not exist in unimodal reasoning.

    For instance:

    • The model incorrectly and confidently explains the presence of an object, while it misidentifies the object.
    • The model erroneously alternates between the verbal and visual texts. The image may show 2020 at a text which states 2019.
    • The model over-relies on one input, disregarding that the other relevant input may be more informative.
    • In unimodal systems, failure is more detectable. As an instance, the text model may generate a permissive false text.
    • Anomalies like these can double in cross-modal systems, where the model could misrepresent the text, the image, or the connection between them.

    The reasoning chain, explaining, and debugging are harder for enterprise application.

    3. Demand for Enhancing Quality of Training Data, and More Effort in Data Curation

    Unimodal datasets, either pure text or images, are big, fascinatingly easy to acquire. Multimodal datasets, though, are not only smaller but also require more stringent alignment of different types of data.

    You have to make sure that the following data is aligned:

    • The caption on the image is correct.
    • The transcript aligns with the audio.
    • The bounding boxes or segmentation masks are accurate.
    • The video has a stable temporal structure.

    That means for businesses:

    • More manual curation.
    • Higher costs for labeling.
    • More domain expertise is required, like radiologists for medical imaging and clinical notes.

    The model depends greatly on the data alignment of the cross-modal model.

    4. Complexity of Assessment Along with Richer Understanding

    It is simple to evaluate a model that is unimodal, for example, you could check for precision, recall, BLEU score, or evaluate by simple accuracy. Multimodal reasoning is more difficult:

    • Does the model have accurate comprehension of the image?
    • Does it refer to the right section of the image for its text?
    • Does it use the right language to describe and account for the visual evidence?
    • Does it filter out irrelevant visual noise?
    • Can it keep spatial relations in mind?

    The need for new, modality-specific benchmarks generates further costs and delays in rolling out systems.

    In regulated fields, this is particularly challenging. How can you be sure a model rightly interprets medical images, safety documents, financial graphs, or identity documents?

    5. More Flexibility Equals More Engineering Dependencies

    To build cross-modal architectures, you also need the following:

    • Vision encoder.
    • Text encoder.
    • Audio encoder (if necessary).
    • Multi-head fused attention.
    • Joint representation space.
    • Multimodal runtime optimizers.

    This raises the complexity in engineering:

    • More components to upkeep.
    • More model parameters to control.
    • More pipelines for data flows to and from the model.

    Greater risk of disruptions from failures, like images not loading and causing invalid reasoning.

    In production systems, these dependencies need:

    • More robust CI/CD testing.
    • Multimodal observability.
    • More comprehensive observability practices.
    • Greater restrictions on file uploads for security.

    6. More Advanced Functionality Equals Less Control Over the Model

    Cross-modal models are often “smarter,” but can also be:

    • More likely to give what is called hallucinations, or fabricated, nonsensical responses.
    • More responsive to input manipulations, like modified images or misleading charts.
    • Less easy to constrain with basic controls.

    For example, you might be able to limit a text model by engineering complex prompt chains or by fine-tuning the model on a narrow data set.But machine-learning models can be easily baited with slight modifications to images.

    To counter this, several defenses must be employed, including:

    • Input sanitization.
    • Checking for neural watermarks
    • Anomaly detection in the vision system
    • Output controls based on policy
    • Red teaming for multiple modal attacks.
    • Safety becomes more difficult as the risk profile becomes more detailed.
    • Cross-Modal Intelligence, Higher Value but Slower to Roll Out

    The bottom line with respect to risk is simpler but still real:

    The vision system must be able to perform a wider variety of tasks with greater complexity, in a more human-like fashion while accepting that the system will also be more expensive to build, more expensive to run, and will increasing complexity to oversee from a governance standpoint.

    Cross-modal models deliver:

    • Document understanding
    • PDF and data table knowledge
    • Visual data analysis
    • Clinical reasoning with medical images and notes
    • Understanding of product catalogs
    • Participation in workflow automation
    • Voice interaction and video genera

    Building such models entails:

    • Stronger infrastructure
    • Stronger model control
    • Increased operational cost
    • Increased number of model runs
    • Increased complexity of the risk profile

    Increased value balanced by higher risk may be a fair trade-off.

    Humanized summary

    Cross modal reasoning is the point at which AI can be said to have multiple senses. It is more powerful and human-like at performing tasks but also requires greater resources to operate seamlessly and efficiently. Where data control and governance for the system will need to be more precise.

    The trade-off is more complex, but the end product is a greater intelligence for the system.

    See less
daniyasiddiquiEditor’s Choice
Asked: 07/11/2025In:

agentic AI differ from traditional ML ...

agentic-aiagentsaiartificial intelligenceautonomous-systemsmachine learning
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    An AI agent is But that is not all: An agent is something more than a predictive or classification model; rather, it is an autonomous system that may take an action directed towards some goal. Put differently, An AI agent processes information, but it doesn't stop there. It's in the comprehension, tRead more

    An AI agent is

    But that is not all: An agent is something more than a predictive or classification model; rather, it is an autonomous system that may take an action directed towards some goal.

    Put differently,

    An AI agent processes information, but it doesn’t stop there. It’s in the comprehension, the memory, and the goals that will determine what comes next.

    Let’s consider three key capabilities of an AI agent:

    • Perception: It collects information from sensors, APIs, documents, user prompts, amongst others.
    • Reasoning: It knows context, and it plans or decides what to do next.
    • What it does: Performs an action list; this can invoke another API, write to a file, send an email, or initiate a workflow.

    A classical ML model could predict whether a transaction is fraudulent.

    But an AI agent could:

    • Detect suspicious transactions,
    • Look up the customer’s account history.
    • Send a confirmation email,

    Suspend the account if no response comes and do all that without a human telling it step by step.

    Under the Hood: What Makes an AI Agent “Agentic”?

    Genuinely agentic AI systems, by contrast, extend large language models like GPT-5 or Claude with more layers of processing and give them a much greater degree of autonomy and goal-directedness:

    Goal Orientation:

    • Instead of answering to one prompt, their focus is on an outcome: “book a ticket,” “generate a report”, or “solve a support ticket.”

    Planning and Reasoning:

    • They split a big problem up into smaller steps, for example, “first fetch data, then clean it, then summarize it”.

    Tool Use / API Integration:

    • They can call other functions and APIs. For instance, they could query a database, send an email, or interface to some other system.

    Memory:

    • They remember previous interactions or actions such that multi-turn reasoning and continuity can be achieved.

    Feedback Loops:

    • They can evaluate if they succeeded with their action, or failed, and thus adjust the next action just as human beings do.

    These components make the AI agents feel much less like “smart calculators” and more like “junior digital coworkers”.

    A Practical Example

    Now, let us consider a simple use case comparison wherein health-scheme claim analysis is close to your domain:

    In essence, any regular ML model would take the claims data as input and predict:

    → “The chance of this claim being fraudulent is 82%.”

    An AI agent could:

    • Check the claim.
    • Pull histories of hospitals and beneficiaries from APIs.
    • Check for consistency in the document.
    • Flag the anomalies and give a summary report to an officer.
    • If no response, follow up in 48 hours.

    That is the key shift: the model informs, while the agent initiates.

    Why the Shift to Agentic AI Matters

    Autonomy → Efficiency:

    • Agents can handle a repetitive workflow without constant human supervision.

    Scalability → Real-World Value:

    • You can deploy thousands of agents for customer support, logistics, data validation, or research tasks.

    Context Retention → Better Reasoning:

    • Since they retain memory and context, they can perform multitask processes with ease, much like any human analyst.

    Interoperability → System Integration:

    • They can interact with enterprise systems such as databases, CRMs, dashboards, or APIs to close the gap between AI predictions and business actions.

     Limitations & Ethical Considerations

    While agentic AI is powerful, it has also opened several new challenges:

    • Hallucination risk: agents may act on false assumptions.
    • Accountability: Who is responsible in case an AI agent made the wrong decision?
    • Security: API access granted to agents could be misused and cause damage.
    • Over-autonomy: Many applications, such as those in healthcare or finance.

    do need human-in-the-loop. Hence, the current trend is hybrid autonomy: AI agents that act independently but always escalate key decisions to humans.

    Body Language by Jane Smith

    “An AI agent is an intelligent system that analyzes data while independently taking autonomous actions toward a goal. Unlike traditional ML models that stop at prediction, agentic AI is able to reason, plan, use tools, and remember context effectively bridging the gap between intelligence and action. While the traditional models are static and task-specific, the agentic systems are dynamic and adaptive, capable of handling end-to-end workflows with minimal supervision.”

    See less
daniyasiddiquiEditor’s Choice
Asked: 07/11/2025In:

CNN vs an RNN vs a transformer

cnndeep learningmachine learningneural-networksrnntransformers
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    Understanding the Core Differences That is, by choosing between CNNs, RNNs, and Transformers, you are choosing how a model sees patterns in data: whether they are spatial, temporal, or contextual relationships across long sequences. Let's break that down: 1. Convolutional Neural Networks (CNNs) – BeRead more

    Understanding the Core Differences

    That is, by choosing between CNNs, RNNs, and Transformers, you are choosing how a model sees patterns in data: whether they are spatial, temporal, or contextual relationships across long sequences.

    Let’s break that down:

    1. Convolutional Neural Networks (CNNs) – Best for spatial or grid-like data

    When to use:

    • Use a CNN when your data has a clear spatial structure, meaning that patterns depend on local neighborhoods.
    • Think images, videos, medical scans, satellite imagery, or even feature maps extracted from sensors.

    Why it works:

    • Convolutions used by CNNs are sliding filters that detect local features: edges, corners, colors.
    • As data passes through layers, the model builds up hierarchical feature representations from edges → textures → objects → scenes.

    Example use cases:

    • Image classification (e.g., diagnosing pneumonia from chest X-rays)

    • Object detection (e.g., identifying road signs in self-driving cars)

    • Facial recognition, medical segmentation, or anomaly detection in dashboards

    • Even some analysis of audio spectrograms-a way of viewing sound as a 2D map of frequencies in

    In short: It’s when “where something appears” is more crucial than “when it does.”

    2. Recurrent Neural Networks (RNNs) – Best for sequential or time-series data

    When to use:

    • Use RNNs when order and temporal dependencies are important; current input depends on what has come before.

    Why it works:

    • RNNs have a persistent hidden state that gets updated at every step, which lets them “remember” previous inputs.
    • Variants include LSTM and GRU, which allow for longer dependencies to be captured and avoid vanishing gradients.

    Example use cases:

    • Natural language tasks like Sentiment Analysis, machine translation before transformers took over
    • Time-series forecasting: stock prices, patient vitals, weather data, etc.
    • Sequential data modeling: for example, monitoring hospital patients, ECG readings, anomaly detection in IoT streams.
    • Speech recognition or predictive text

    In other words: RNNs are great when “sequence and timing” is most important – you’re modeling how it unfolds.

    3. Transformers – Best for context-heavy data with long-range dependencies

    When to use:

    • Transformers are currently the state of the art for nearly every task that requires modeling complicated relationships on long sequences-text, images, audio, even structured data.

    Why it works:

    • Unlike RNNs, which process data one step at a time, transformers make use of self-attention — a mechanism that allows the model to look at all parts of the input at once and decide which parts are most relevant to each other.

    This gives transformers three big advantages:

    • Parallelization: Training is way faster because inputs are processed simultaneously.
    • Long-range understanding: They are global in capturing dependencies, for example, word 1 affecting word 100.
    • Adaptability: Works across multiple modalities, such as text, images, code, etc.

    Example use cases:

    • NLP: ChatGPT, BERT, T5, etc.
    • Vision: The ViT now competes with the CNN for image recognition.
    • Audio/Video: Speech-to-text, music generation, multimodal tasks.
    • Health & business: Predictive analytics using structured plus unstructured data such as clinical notes and sensor data.

    In other words, Transformers are ideal when global context and scalability are critical — when you need the model to understand relationships anywhere in the sequence.

     Example Analogy (for Human Touch)

    Imagine you are analyzing a film:

    • A CNN focuses on every frame; the visuals, the color patterns, who’s where on screen.
    • An RNN focuses on how scenes flow over time the storyline, one moment leading to another.
    • A Transformer reads the whole script at once: character relationships, themes, and how the ending relates to the beginning.

    So, it depends on whether you are analyzing visuals, sequence, or context.

    Summary Answer for an Interview

    I will choose a CNN if my data is spatially correlated, such as images or medical scans, since it does a better job of modeling local features. But if there is some strong temporal dependence in my data, such as time-series or language, I will select an RNN or an LSTM, which does the processing sequentially. If the task, however, calls for an understanding of long-range dependencies or relationships, especially for large and complex datasets, then I would use a Transformer. Recently, Transformers have generalized across vision, text, and audio and therefore have become the default solution for most recent deep learning applications.

    See less
daniyasiddiquiEditor’s Choice
Asked: 19/10/2025In:

AI model to use (for a given task)

ai model selectiondeep learningmachine learningmodel choicemodel performancetask-specific models
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    1. Start with the Problem — Not the Model Specify what you actually require even before you look at models. Ask yourself: What am I trying to do — classify, predict, generate content, recommend, or reason? What is the input and output we have — text, images, numbers, sound, or more than one (multimoRead more

    1. Start with the Problem — Not the Model

    Specify what you actually require even before you look at models.

    Ask yourself:

    • What am I trying to do — classify, predict, generate content, recommend, or reason?
    • What is the input and output we have — text, images, numbers, sound, or more than one (multimodal)?
    • How accurate or original should the system be?

    For example:

    • If you want to summarize patient reports → use a large language model (LLM) fine-tuned for summarization.
    • If you want to diagnose pneumonia on X-rays → use a vision model fine-tuned on medical images (e.g., EfficientNet or ViT).
    • If you want to answer business questions in natural language → use a reasoning model like GPT-4, Claude 3, or Gemini 1.5.

    When you are aware of the task type, you’ve already completed half the job.

     2. Match the Model Type to the Task

    With this information, you can narrow it down:

    Task Type\tModel Family\tExample Models
    Text generation / summarization\tLarge Language Models (LLMs)\tGPT-4, Claude 3, Gemini 1.5
    Image generation\tDiffusion / Transformer-based\tDALL-E 3, Stable Diffusion, Midjourney
    Speech to text\tASR (Automatic Speech Recognition)\tWhisper, Deepgram
    Text to speech\tTTS (Text-to-Speech)\tElevenLabs, Play.ht
    Image recognition\tCNNs / Vision Transformers\tEfficientNet, ResNet, ViT
    Multi-modal reasoning
    Unified multimodal transformers
    GPT-4o, Gemini 1.5 Pro
    Recommendation / personalization
    Collaborative filtering, Graph Neural Nets
    DeepFM, GraphSage

    If your app uses modalities combined (like text + image), multimodal models are the way to go.

     3. Consider Scale, Cost, and Latency

    Not every problem requires a 500-billion-parameter model.

    Ask:

    • Do I require state-of-the-art accuracy or good-enough speed?
    • How much am I willing to pay per query or per inference?

    Example:

    • Customer support chatbots → smaller, lower-cost models like GPT-3.5, Llama 3 8B, or Mistral 7B.
    • Scientific reasoning or code writing → larger models like GPT-4-Turbo or Claude 3 Opus.
    • On-device AI (like in mobile apps) → quantized or distilled models (Gemma 2, Phi-3, Llama 3 Instruct).

    The rule of thumb:

    • “Use the smallest model that’s good enough for your use case.”
    • This is budget-friendly and makes systems responsive.

     4. Evaluate Data Privacy and Deployment Needs

    • Your data is sensitive (health, finance, government), and you want to control where and how the model runs.
    • Cloud-hosted proprietary models (e.g., GPT-4, Gemini) give excellent performance but little data control.
    • Self-hosted or open-source models (e.g., Llama 3, Mistral, Falcon) can be securely deployed on your servers.

    If your business requires ABDM/HIPAA/GDPR compliance, self-hosting or API use of models is generally the preferred option.

     5. Verify on Actual Data

    The benchmark score of a model does not ensure it will work best for your data.
    Always pilot test it on a very small pilot dataset or pilot task first.

    Measure:

    • Accuracy or relevance (depending on task)
    • Speed and cost per request
    • Robustness (does it crash on hard inputs?)
    • Bias or fairness (any demographic bias?)

    Sometimes a little fine-tuned model trumps a giant general one because it “knows your data better.”

    6. Contrast “Reasoning Depth” with “Knowledge Breadth”

    Some models are great reasoners (they can perform deep logic chains), while others are good knowledge retrievers (they recall facts quickly).

    Example:

    • Reasoning-intensive tasks: GPT-4, Claude 3 Opus, Gemini 1.5 Pro
    • Knowledge-based Q&A or embeddings: Llama 3 70B, Mistral Large, Cohere R+

    If your task concerns step-by-step reasoning (such as medical diagnosis or legal examination), use reasoning models.

    If it’s a matter of getting information back quickly, retrieval-augmented smaller models could be a better option.

     7. Think Integration & Tooling

    Your chosen model will have to integrate with your tech stack.

    Ask:

    • Does it support an easy API or SDK?
    • Will it integrate with your existing stack (React, Node.js, Laravel, Python)?
    • Does it support plug-ins or direct function call?

    If you plan to deploy AI-driven workflows or microservices, choose models that are API-friendly, reliable, and provide consistent availability.

     8. Try and Refine

    No choice is irreversible. The AI landscape evolves rapidly — every month, there are new models.

    A good practice is to:

    • Start with a baseline (e.g., GPT-3.5 or Llama 3 8B).
    • Collect performance and feedback metrics.
    • Scale up to more powerful or more specialized models as needed.
    • Have fall-back logic — i.e., if one API will not do, another can take over.

    In Short: Selecting the Right Model Is Selecting the Right Tool

    It’s technical fit, pragmatism, and ethics.

    Don’t go for the biggest model; go for the most stable, economical, and appropriate one for your application.

    “A great AI product is not about leveraging the latest model — it’s about making the best decision with the model that works for your users, your data, and your purpose.”

    See less
daniyasiddiquiEditor’s Choice
Asked: 13/10/2025In:

AI

aiartificial intelligenceautomationfuture-of-techmachine learningtechnology
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    1. The Simple Idea: Machines Taught to "Think" Artificial Intelligence is the design of making computers perform intelligent things — not just by following instructions, but actually learning from information and improving with time. In regular programming, humans teach computers to accomplish thingRead more

    1. The Simple Idea: Machines Taught to “Think”

    Artificial Intelligence is the design of making computers perform intelligent things — not just by following instructions, but actually learning from information and improving with time.

    In regular programming, humans teach computers to accomplish things step by step.

    In AI, computers learn to resolve things on their own by gaining expertise on patterns in information.

    For example

    When Siri quotes back the weather to you, it is not reading from a script. It is recognizing your voice, interpreting your question, accessing the right information, and responding in its own words — all driven by AI.

    2. How AI “Learns” — The Power of Data and Algorithms

    Computers are instructed with so-called machine learning —inferring catalogs of vast amounts of data so that they may learn patterns.

    • Machine Learning (ML): The machine learns by example, not by rule. Display a thousand images of dogs and cats, and it may learn to tell them apart without learning to do so.
    • Deep Learning: Latest generation of ML based on neural networks —stacks of algorithms imitating the way we think.

    That’s how machines can now identify faces, translate text, or compose music.

    3. Examples of AI in Your Daily Life

    You probably interact with AI dozens of times a day — maybe without even realizing it.

    • Your phone: Face ID, voice assistants, and autocorrect.
    • Streaming: Netflix or Spotify recommends you like something.
    • Shopping: Amazon’s “Recommended for you” page.
    • Health care: AI is diagnosing diseases from X-rays faster than doctors.
    • Cars: Self-driving vehicles with sensors and AI delivering split-second decisions.

    AI isn’t science fiction anymore — it’s present in our reality.

     4. AI types

    AI isn’t one entity — there are levels:

    • Narrow AI (Weak AI): Designed to perform a single task, like ChatGPT responding or Google Maps route navigation.
    • General AI (Strong AI): A Hypothetical kind that would perhaps understand and reason in several fields as any common human individual, yet to be achieved.
    • Superintelligent AI: Another level higher than human intelligence — still a future goal, but widely seen in the movies.

    We already have Narrow AI, mostly, but it is already incredibly powerful.

     5. The Human Side — Pros and Cons

    AI is full of promise and also challenges our minds to do the hard thinking.

    Advantages:

    • Smart healthcare diagnosis
    • Personalized learning
    • Weather prediction and disaster simulations
    • Faster science and technology innovation

    Disadvantages:

    • Bias: AI can be biased in decision-making if AI is trained using biased data.
    • Job loss: Automation will displace some jobs, especially repetitive ones.
    • Privacy: AI systems gather huge amounts of personal data.
    • Ethics: Who would be liable if an AI erred — the maker, the user, or the machine?

    The emergence of AI presses us to redefine what it means to be human in an intelligent machine-shared world.

    6. The Future of AI — Collaboration, Not Competition

    The future of AI is not one of machines becoming human, but humans and AI cooperating. Consider physicians making diagnoses earlier with AI technology, educators adapting lessons to each student, or cities becoming intelligent and green with AI planning.

    AI will progress, yet it will never cease needing human imagination, empathy, and morals to steer it.

     Last Thought

    Artificial Intelligence is not a technology — it’s a demonstration of humans of the necessity to understand intelligence itself. It’s a matter of projecting our minds beyond biology. The more we advance in AI, the more the question shifts from “What can AI do?” to “How do we use it well to empower all?”

    See less
daniyasiddiquiEditor’s Choice
Asked: 25/09/2025In: ,

aligning large language models with h ...

ai ecosystemfalconlanguage-modelsllamamachine learningmistralopen-source
  1. daniyasiddiqui
    daniyasiddiqui Editor’s Choice

    What “Aligning with Human Values” Means Before we dive into the methods, a quick refresher: when we say “alignment,” we mean making LLMs behave in ways that are consistent with what people value—that includes fairness, honesty, helpfulness, respecting privacy, avoiding harm, cultural sensitivity, etRead more

    What “Aligning with Human Values” Means

    Before we dive into the methods, a quick refresher: when we say “alignment,” we mean making LLMs behave in ways that are consistent with what people value—that includes fairness, honesty, helpfulness, respecting privacy, avoiding harm, cultural sensitivity, etc. Because human values are complex, varied, sometimes conflicting, alignment is more than just “don’t lie” or “be nice.”

    New / Emerging Methods in HLM Alignment

    Here are several newer or more refined approaches researchers are developing to better align LLMs with human values.

    1. Pareto Multi‑Objective Alignment (PAMA)

    • What it is: Most alignment methods optimize for a single reward (e.g. “helpfulness,” or “harmlessness”). PAMA is about balancing multiple objectives simultaneously—like maybe you want a model to be informative and concise, or helpful and creative, or helpful and safe.
    • How it works: It transforms the multi‑objective optimization (MOO) problem into something computationally tractable (i.e. efficient), finding a “Pareto stationary point” (a state where you can’t improve one objective without hurting another) in a way that scales well.
    • Why it matters: Because real human values often pull in different directions. A model that, say, always puts safety first might become overly cautious or bland, and one that is always expressive might sometimes be unsafe. Finding trade‑offs explicitly helps.

    2. PluralLLM: Federated Preference Learning for Diverse Values

    • What it is: A method to learn what different user groups prefer without forcing everyone into one “average” view. It uses federated learning so that preference data stays local (e.g., with a community or user group), doesn’t compromise privacy, and still contributes to building a reward model.
    • How it works: Each group provides feedback (or preferences). These are aggregated via federated averaging. The model then aligns to those aggregated preferences, but because the data is federated, groups’ privacy is preserved. The result is better alignment to diverse value profiles.
    • Why it matters: Human values are not monoliths. What’s “helpful” or “harmless” might differ across cultures, age groups, or contexts. This method helps LLMs better respect and reflect that diversity, rather than pushing everything to a “mean” that might misrepresent many.

    3. MVPBench: Global / Demographic‑Aware Alignment Benchmark + Fine‑Tuning Framework

    • What it is: A new benchmark (called MVPBench) that tries to measure how well models align with human value preferences across different countries, cultures, and demographics. It also explores fine‑tuning techniques that can improve alignment globally.
    • Key insights: Many existing alignment evaluations are biased toward a few regions (English‑speaking, WEIRD societies). MVPBench finds that models often perform unevenly: aligned well for some demographics, but poorly for others. It also shows that lighter fine‑tuning (e.g., methods like LoRA, Direct Preference Optimization) can help reduce these disparities.
    • Why it matters: If alignment only serves some parts of the world (or some groups within a society), the rest are left with models that may misinterpret or violate their values, or be unintentionally biased. Global alignment is critical for fairness and trust.

    4. Self‑Alignment via Social Scene Simulation (“MATRIX”)

    • What it is: A technique where the model itself simulates “social scenes” or multiple roles around an input query (like imagining different perspectives) before responding. This helps the model “think ahead” about consequences, conflicts, or values it might need to respect.
    • How it works: You fine‑tune using data generated by those simulations. For example, given a query, the model might role play as user, bystander, potential victim, etc., to see how different responses affect those roles. Then it adjusts. The idea is that this helps it reason about values in a more human‑like social context.
    • Why it matters: Many ethical failures of AI happen not because it doesn’t know a rule, but because it didn’t anticipate how its answer would impact people. Social simulation helps with that foresight.

    5. Causal Perspective & Value Graphs, SAE Steering, Role‑Based Prompting

    • What it is: Recent work has started modeling how values relate to each other inside LLMs — i.e. building “causal value graphs.” Then using those to steer models more precisely. Also using methods like sparse autoencoder steering and role‑based prompts.

    How it works:
    • First, you estimate or infer a structure of values (which values influence or correlate with others).
    • Then, steering methods like sparse autoencoders (which can adjust internal representations) or role‑based prompts (telling the model to “be a judge,” “be a parent,” etc.) help shift outputs in directions consistent with a chosen value.

    • Why it matters: Because sometimes alignment fails due to hidden or implicit trade‑offs among values. For example, trying to maximize “honesty” could degrade “politeness,” or “transparency” could clash with “privacy.” If you know how values relate causally, you can more carefully balance these trade‑offs.

    6. Self‑Alignment for Cultural Values via In‑Context Learning

    • What it is: A simpler‑but‑powerful method: using in‑context examples that reflect cultural value statements (e.g. survey data like the World Values Survey) to “nudge” the model at inference time to produce responses more aligned with the cultural values of a region.
    • How it works: You prepare some demonstration examples that show how people from a culture responded to value‑oriented questions; then when interacting, you show those to the LLM so it “adopts” the relevant value profile. This doesn’t require heavy retraining.
    • Why it matters: It’s a relatively lightweight, flexible method, good for adaptation and localization without needing huge data/fine‑tuning. For example, responses in India might better reflect local norms; in Japan differently etc. It’s a way of personalizing / contextualizing alignment.

    Trade-Offs, Challenges, and Limitations (Human Side)

    All these methods are promising, but they aren’t magic. Here are where things get complicated in practice, and why alignment remains an ongoing project.

    • Conflicting values / trade‑offs: Sometimes what one group values may conflict with what another group values. For instance, “freedom of expression” vs “avoiding offense.” Multi‑objective alignment helps, but choosing the balance is inherently normative (someone must decide).
    • Value drift & unforeseen scenarios: Models may behave well in tested cases, but fail in rare, adversarial, or novel situations. Humans don’t foresee everything, so there’ll always be gaps.
    • Bias in training / feedback data: If preference data, survey data, cultural probes are skewed toward certain demographics, the alignment will reflect those biases. It might “over‑fit” to values of some groups, under‑represent others.
    • Interpretability & transparency: You want reasons why the model made certain trade‑offs or gave a certain answer. Methods like causal value graphs help, but much of model internal behavior remains opaque.
    • Cost & scalability: Some methods require more data, more human evaluators, or more compute (e.g. social simulation is expensive). Getting reliable human feedback globally is hard.
    • Cultural nuance & localization: Methods that work in one culture may fail or even harm in another, if not adapted. There’s no universal “values” model.

    Why These New Methods Are Meaningful (Human Perspective)

    Putting it all together: what difference do these advances make for people using or living with AI?

    • For everyday users: better predictability. Less likelihood of weird, culturally tone‑deaf, or insensitive responses. More chance the AI will “get you” — in your culture, your language, your norms.
    • For marginalized groups: more voice in how AI is shaped. Methods like pluralistic alignment mean you aren’t just getting “what the dominant culture expects.”
    • For build‑and‑use organizations (companies, developers): more tools to adjust models for local markets or special domains without starting from scratch. More ability to audit, test, and steer behavior.
    • For society: less risk of AI reinforcing biases, spreading harmful stereotypes, or misbehaving in unintended ways. More alignment can help build trust, reduce harms, and make AI more of a force for good.
    See less
mohdanasMost Helpful
Asked: 22/09/2025In:

it different from regular AI models

ai technology deep learningartificial intelligencedeep learningmachine learningmultimodal ai
  1. mohdanas
    mohdanas Most Helpful

    What is Multimodal AI? In its simplest definition, multimodal AI is a form of artificial intelligence that can comprehend and deal with more than one kind of input—at least text, images, audio, and even video—simultaneously. Consider how humans communicate: when you're talking with a friend, you donRead more

    What is Multimodal AI?

    In its simplest definition, multimodal AI is a form of artificial intelligence that can comprehend and deal with more than one kind of input—at least text, images, audio, and even video—simultaneously.

    Consider how humans communicate: when you’re talking with a friend, you don’t solely depend on language. You read facial expressions, tone of voice, and body language as well. That’s multimodal communication. Multimodal AI is attempting to do the same—soaking up and linking together different channels of information to better understand the world.

    How is it Different from Regular AI Models?

    kind of traditional or “single-modal” AI models are typically trained to process only one :

    • A text-based model such as vintage chatbots or search engines can process only written language.
    • An image recognition model can recognize cats in pictures but can’t explain them in words.
    • A speech-to-text model can convert audio into words, but it won’t also interpret the meaning of what was said in relation to an image or a video.
    • Multimodal AI turns this limitation on its head. Rather than being tied to a single ability, it learns across modalities. For instance:
    • You upload an image of your fridge, and the AI not only identifies the ingredients but also provides a text recipe suggestion.
    • You play a brief clip of a soccer game, and it can describe the action along with summarizing the play-by-play.

    You say a question aloud, and it not only hears you but also calls up similar images, diagrams, or text to respond.

     Why Does it Matter for Humans?

    • Multimodal AI seems like a giant step forward because it gets closer to the way we naturally think and learn.
    • A kid discovers that “dog” is not merely a word—they hear someone say it, see the creature, touch its fur, and integrate all those perceptions into one idea.
    • Likewise, multimodal AI can ingest text, pictures, and sounds, and create a richer, more multidimensional understanding.

    More natural, human-like conversations. Rather than jumping between a text app, an image app, and a voice assistant, you might have one AI that does it all in a smooth, seamless way.

     Opportunities and Challenges

    • Opportunities: Smarter personal assistants, more accessible technology (assisting people with disabilities through the marriage of speech, vision, and text), education breakthroughs (visual + verbal instruction), and creative tools (using sketches to create stories or songs).
    • Challenges: Building models for multiple types of data takes enormous computing resources and concerns privacy—because the AI is not only consuming your words, it might also be scanning your images, videos, or even voice tone. There’s also a possibility that AI will commit “multimodal mistakes”—such as misinterpreting sarcasm in talk or overreading an image.

     In Simple Terms

    If standard AI is a person who can just read books but not view images or hear music, then multimodal AI is a person who can read, watch, listen, and then integrate all that knowledge into a single greater, more human form of understanding.

    It’s not necessarily smarter—it’s more like how we sense the world.

    See less