Before my eyes are fully open, my hand is already moving.
Not toward water.
Not toward the window.
Toward you.
One small screen.
One quick check.
Just to see what happened while I was asleep.
—
There is a quiet moment right after waking when the mind is still soft.
No plans have claimed it yet.
No messages have shaped it.
No comparisons have entered.
That moment used to belong to silence.
Now it belongs to whatever is waiting on the other side of the lock screen.
—
I tell myself it takes only a few seconds.
But those seconds set the tone for the next hour.
Someone else’s urgency becomes my first thought.
Someone else’s highlight becomes my first feeling.
Someone else’s day arrives before my own has even started.
—
What would happen if I waited ten minutes?
If I let the room stay quiet.
If I let my own thoughts arrive first.
Would the day feel different?
Or would I simply notice how automatic the reach has become?
—
The phone is not the problem.
The habit of offering it the first piece of my attention is.
That first piece is the one that shapes everything that follows.
—
Tomorrow morning, I will try to leave you on the table for a little longer.
Just long enough to remember what my own morning feels like before anyone else has entered it.
A pause in the queue.
A gap between messages.
A few quiet seconds while the kettle boils.
And almost immediately I reach for something to put inside you.
A scroll. A check. A quick reply that could have waited.
Because an empty moment feels like wasted time.
Or worse — like I might have to meet myself in it.
—
When did rest start needing a reason?
Children fall into empty moments naturally.
They watch a bug cross the floor.
They stare out a window without calling it productive.
Somewhere we learned that every second should be filled with progress, entertainment, or proof that we are not falling behind.
—
So we keep the feed open.
We open another tab.
We start a conversation just so the silence does not last long enough to ask a question.
—
But empty moments are not empty.
They are the only places left where the mind can finish a thought it started hours ago.
Where a feeling can settle instead of being pushed aside.
Where the next right thing becomes visible only because nothing else is shouting.
—
Perhaps the habit is not the problem.
Perhaps the fear is.
The fear that if we stop filling the space, we might discover we are already enough without the constant motion.
—
So next time an empty moment appears, maybe try leaving it alone for thirty seconds.
Just to see what shows up when nothing is forced to.
—
Sometimes the most useful thing you can do is nothing at all.
A visual metaphor for extra inference effort, alternative solutions and checking—not a guarantee of correctness.
Reasoning models are AI models trained or adapted to handle tasks that benefit from several connected steps. Test-time compute is the computing effort used while solving a request, after training. Extra inference effort can support longer solution attempts, alternative candidates or verification. It does not guarantee a correct answer.
These study notes explain the terms, main methods, research evidence and practical trade-offs. Examples are illustrative, not results from a SANSANI product test.
AI explained — article 3 of 11. Updated 15 November 2026.
Quick notes: the whole topic in six points
Training compute: effort spent learning or adjusting a model’s parameters.
Test-time compute: effort spent producing and checking a response during inference.
Reasoning model: a practical label for models optimised for multi-step problem solving; not a promise of human-like thought.
More effort can help: especially when a task has useful intermediate steps and checkable outcomes.
More effort can be wasted: repeating an incorrect assumption does not create missing evidence.
Evaluate the result: compare correctness, total cost and response time—not answer length alone.
1. What counts as a reasoning task?
A task needs reasoning when the answer depends on connecting facts, applying constraints or checking consequences. It can also require retrieval, calculation or domain knowledge.
Simple transformation: turn a supplied paragraph into bullet points.
Constraint problem: arrange deliveries without exceeding vehicle capacity or missing time windows.
Debugging: find why a program fails, propose a correction and test whether the failure is resolved.
Evidence comparison: identify contradictions across documents and separate supported conclusions from assumptions.
Remember: a difficult question is not always a reasoning problem. If the answer depends on a fact absent from the input and model knowledge, additional computation alone cannot reliably supply that fact. Retrieval or a clarifying question may be necessary.
2. Training compute vs test-time compute
Aspect
Training
Inference / test time
When?
During model development or adaptation
While responding to a request
Main purpose
Learn or adjust behaviour through parameter updates
Use the trained model to solve the current task
Possible extra effort
Additional optimisation or training data
Longer attempts, more candidates, checks or tool use
Persistent learning?
Parameter updates can persist
Ordinary extra inference does not itself retrain the model
Analogy: training resembles preparation before an exam; test-time compute resembles the effort spent on an exam question. This is only an analogy. Neural networks do not learn or think exactly as people do.
Snell and colleagues’ 2024 study examined how to allocate inference computation. It found that the effectiveness of different methods depended on problem difficulty, supporting adaptive allocation rather than treating every request identically. Its comparisons were made under particular experimental conditions, not a universal guarantee that a smaller model with more time beats a larger model. [1]
3. Where the extra computation can go
A. A longer sequential attempt
Spend more generation steps developing, revising or checking a solution.
Potential benefit: discover an overlooked constraint or correct an intermediate mistake.
Trade-off: extra latency and computation; a longer attempt can still repeat the same mistake.
The 2025 s1 paper explored a trained model with “budget forcing” to control inference effort, including extending generation when it tried to stop. It reported improvements in its evaluated setting. This was a specific training-and-inference method, not evidence that adding “wait” to any ordinary prompt reliably improves any model. [2]
B. Several candidate solutions
Generate multiple attempts instead of relying on the first one.
Select through a scoring method, answer agreement or an external check.
Potential benefit: explore alternatives.
Trade-off: selection is difficult when several candidates share a false assumption.
Self-consistency research samples different solution paths and aggregates their answers. Its original work reported gains on arithmetic and commonsense benchmarks. Agreement is a useful signal in some settings, but several agreeing outputs are not equivalent to independent evidence. [3]
C. Search and verification
Explore possible solution steps and evaluate which are promising.
A verifier can score a final answer or intermediate steps.
A calculator, test suite or constraint checker can sometimes provide a firmer check than another generated explanation.
A flawed verifier may reward plausible errors, so verification quality matters.
Let’s Verify Step by Step compared feedback on final outcomes with feedback on intermediate steps when training mathematical verifiers. The paper found advantages for process supervision in its MATH experiments. That is evidence for the studied approach, not proof that every visible step in any model’s answer has been verified. [4]
4. How training and inference work together
Giving a model more computation and training it to use that computation effectively are related but distinct decisions.
Supervised fine-tuning: learn from examples of desired responses or solution behaviour.
Reinforcement learning: adjust behaviour using rewards associated with generated responses.
Inference strategy: decide how much effort to spend and how to select or check outputs at request time.
The DeepSeek-R1 research studied reinforcement learning for reasoning. Its R1-Zero experiment applied reinforcement learning without a preliminary supervised fine-tuning stage; the broader R1 work incorporated additional training stages. Keep those experiments distinct. The publication supports reinforcement learning as one route to improved reasoning behaviour, not the claim that every reasoning model is trained in the same way. [5]
Useful distinction: the model’s training influences the strategies it can use. The inference system influences which strategies are attempted, how often and under what budget.
5. Worked example: planning a small delivery run
Hypothetical problem: three deliveries, two vans, one departure time.
Van A holds at most 10 boxes. Van B holds at most 6.
Customer X needs 6 boxes by 10:00.
Customer Y needs 4 boxes by 10:30.
Customer Z needs 6 boxes by 10:15.
Both vans leave the depot at 09:00. Loading is complete.
Assume fixed travel times, immediate unloading and no return-trip requirement for this simplified exercise.
Candidate 1: check the capacity before the timetable
Put X and Y on Van A: 6 + 4 = 10 boxes. Capacity satisfied.
Put Z on Van B: 6 boxes. Capacity satisfied.
Van A reaches X at 09:30 and Y at 09:50. Both deadlines satisfied.
Van B reaches Z at 09:45. Deadline satisfied.
Conclusion: this is a feasible plan under the stated assumptions. It is not a proven cheapest or fastest plan; the problem supplied no complete distance matrix or cost objective.
Candidate 2: an attractive grouping that fails
Put X and Z together on Van A.
It requires 6 + 6 = 12 boxes.
Van A’s limit is 10. Reject the plan before claiming it works.
What useful extra effort changes
First attempt: propose a route.
Check: calculate loads and arrivals.
Revise: reject assignments that violate capacity or deadlines.
Report: present the feasible plan and its assumptions.
The value lies in checking constraints. A long story about efficient logistics would add little if it never calculated the loads.
In a real operation, ask for unloading duration, variable traffic, breaks, vehicle restrictions and the actual optimisation objective. Extra inference cannot justify silently inventing those values.
6. When more reasoning effort is worth trying
Good candidate: a bounded problem with several interacting constraints and a checkable answer.
Good candidate: code changes that can be exercised by meaningful tests.
Good candidate: comparing alternatives where the evaluation criteria are explicit.
Often unnecessary: routine formatting, brief rewriting or a straightforward lookup.
Wrong remedy: missing source material, an ambiguous goal or unavailable current facts. Resolve the information gap first.
Practical rule: begin with the least expensive method that meets the quality requirement. Increase effort when tests show a worthwhile improvement.
7. Limits: longer thinking is not proof
Correctness and explanation are different checks
A convincing explanation can accompany an incorrect answer. Ask for concise assumptions, calculations, references and checkable results. Do not use the length of a displayed reasoning trace as a reliability score.
Anthropic’s April 2025 research found that reasoning traces did not always disclose answer hints that influenced responses in its experiments. The study used specific models and constructed evaluation settings, so it should not be generalised into a claim that every explanation is deceptive. It does show why a generated account is not a guaranteed record of every influence on the answer. [6]
Cost and latency need separate measurement
More candidates can require more total compute even if run simultaneously.
Parallel execution can reduce waiting time without reducing the overall work.
Longer sequential attempts generally add steps before the answer is ready.
Actual charges depend on the service’s current pricing and accounting. No fixed price multiplier applies to every model.
Tools do not make a model infallible
A calculator can perform the wrong calculation if the model supplies the wrong inputs. A test suite can pass while omitting an important requirement. Check the connection between the user’s question, the tool call and the result.
8. A simple evaluation worksheet
Use the same representative problems at two effort levels. Include easy cases, difficult cases and cases that cannot be answered from the supplied information.
Task: what exactly should the model accomplish?
Success condition: which facts, constraints or tests establish correctness?
Baseline: what does a fast, ordinary attempt achieve?
Extra effort: which setting or strategy changes?
Result: does correctness improve, including on previously unseen cases?
Resources: what are total compute cost, waiting time and review effort?
Failure handling: does it acknowledge missing information and stop appropriately?
Illustrative calculation: if a method correctly answers 80 out of 100 evaluation questions and another answers 88, the improvement is 8 percentage points, or 10% relative to the original 80 correct answers. Those figures are made up to explain reporting, not measured model results.
Do not judge from one lucky answer. Record the model version, instructions, available tools and effort settings so the comparison can be repeated.
9. Common questions — short answers
Does test-time compute make the model permanently smarter?
Ordinary inference uses the trained parameters to handle the current request. Extra computation alone does not automatically create a persistent parameter update.
Is a reasoning model an AI agent?
Not necessarily. Reasoning concerns problem solving; an agentic system selects and performs actions through tools. An agent can use a reasoning model. Read the agentic AI notes for the system-level distinction.
Can a reasoning model also be multimodal?
Yes. These describe different capabilities. A system can reason about supported image, text or other inputs. See our multimodal AI guide.
Should every answer show a long chain of thought?
No. A concise explanation with assumptions, evidence and verifiable calculations can be more useful. Displayed text should not be treated as a complete diagnostic record of the model.
Does a slow answer mean the model reasoned more carefully?
No. Waiting can also reflect server load, tool delays or other system behaviour. Evaluate the output and recorded resource use rather than the pause alone.
Revision notes: what to remember
Training builds or adapts capability; inference applies it to a request.
Extra compute can support longer attempts, alternatives and verification.
Task difficulty, model capability and checking quality affect the benefit.
Missing information needs evidence or clarification, not invented certainty.
Choose effort by measured value: correctness, cost, latency and human review.
A conceptual view of combining modalities. Supported inputs, outputs and processing methods vary by system.
Multimodal AI processes or relates more than one kind of information, such as text, images, audio and video. A multimodal system might answer a question about a photograph, connect a spoken explanation to a diagram, or retrieve an image using a written description. Its value comes from using relationships between those inputs, not simply accepting several file extensions.
Suppose a teacher says, “The important change is here,” while pointing to a chart. A transcript preserves the words but not necessarily the point on the chart. A screenshot preserves the chart but not the explanation. A useful multimodal system needs to connect the right words, visual region and moment in time.
This guide explains how that connection can work, what different architectures preserve or discard, and how to evaluate an answer. The worked examples are hypothetical; they are not claims that SANSANI has tested a particular product.
A modality is a type or channel of information. Written language, visual information and acoustic signals can provide different evidence about the same event. Multimodal learning has a long research history: a 2017 survey organised its central challenges into representation, translation, alignment, fusion and co-learning. The field predates today’s conversational assistants. [1]
For a reader, the practical question is which signals a system can actually use. A product may accept a video file but analyse only selected frames. Another may extract its soundtrack, transcribe speech and pass the transcript to a text model. Both can be useful, but they preserve different information.
Input and output capabilities are separate. A system can accept an image and produce only text. A model that generates images from text is also connecting modalities, but it does not automatically have the same strengths as a visual question-answering model. Avoid treating “multimodal” as a promise that every input and output combination is supported.
Common modalities and the evidence they can provide
Input
Useful evidence
Example question
Text
Words, explicit instructions and written facts
What does this paragraph claim?
Image
Visible layout, objects, shapes and marks
Which label belongs to this chart axis?
Audio
Speech and potentially non-speech sounds
What was said, and where does the alarm occur?
Video
Visual change over time, potentially with audio
Which step happens before the object is moved?
The word “potentially” matters. Support for a modality does not establish that a given model can accurately solve every task involving it.
How does multimodal AI connect different inputs?
There is no single architecture used by all multimodal systems. A useful conceptual view is: prepare the inputs, represent them numerically, connect relevant information and produce an output. These stages may be separate components or tightly integrated within a model.
This conceptual diagram simplifies a varied family of architectures. Supporting several modalities does not guarantee accurate interpretation.
Representation: turning signals into useful features
An encoder converts an input into a numerical representation. A vision encoder can represent image information; a language component represents text. Those representations are learned patterns, not a perfect copy of human perception.
CLIP, described in a 2021 research paper, learns relationships between images and text by training on image–text pairs. Its contrastive approach makes matching pairs more similar in a learned representation space than mismatched pairs. That supports tasks such as comparing an image with candidate descriptions. CLIP itself is not a general conversational assistant or image generator. [2]
Alignment: matching the right pieces
Alignment concerns correspondence. A word might relate to an object region, a spoken phrase to a video segment, or a caption to an entire image. Matching “the red line” to the correct curve matters more than recognising that the picture contains a chart.
Consider two speakers discussing two different slides. If the system pairs the first explanation with the second slide, it can produce a fluent but misleading summary. Timing and context are part of the evidence, not merely presentation details.
Fusion and interaction: using information together
Fusion concerns combining information from modalities. Some designs combine intermediate representations; others combine separate predictions. Attention-based connections can allow language processing to use visual features. The right design depends on the task and available data. [1]
Flamingo’s 2022 architecture is a research example that bridges pretrained vision and language models and handles interleaved visual and textual inputs. Its paper evaluates image and video tasks. It illustrates one design approach, rather than a blueprint followed by every current product. [3]
Instruction following: responding to the question
Recognising visual patterns is different from following a user’s request about them. The LLaVA research introduced a vision-language assistant connecting a vision encoder with a large language model and using visual instruction tuning. It is an example of training a system to answer requests involving an image and language together. [4]
For users, the distinction explains why a model can describe a scene generally yet struggle with a precise question about a tiny symbol. Different tasks demand different kinds of training and evidence.
A multimodal model and a tool pipeline are different designs
A combined model can work directly with representations from supported modalities. A modular pipeline can instead convert information into an intermediate form: optical character recognition (OCR) extracts visible text, speech recognition produces a transcript, and a text model processes those results.
Speech-to-text is a substantial AI task in its own right. The Whisper research paper, published in 2022, studied speech recognition using large-scale weak supervision. A transcription component can be valuable inside a wider system without giving the downstream text model direct access to the original sound. [5]
Neither approach is universally better. Evaluate the task’s quality, latency, cost and privacy requirements.
Imagine asking what a speaker said. A transcript pipeline may be enough. Ask whether a warning beep occurred during the demonstration, and a transcript containing only spoken words may omit the necessary evidence. Similarly, OCR may preserve a chart title while losing the relation between a curve and its legend.
Neither architecture wins automatically. Modular stages can make intermediate results easier to inspect and replace. Joint processing can retain relationships that a text-only intermediate representation omits. Both need evaluation on actual tasks; the marketing term “native multimodal” does not substitute for evidence.
Worked example: explaining a recorded lesson
Hypothetical task: a student has permission to use a short recorded lesson containing narration and slides. They ask: “Explain the chart discussed between 02:00 and 02:40. State the labels and values you can read, distinguish observation from interpretation, and tell me what is unclear.”
Suppose the relevant slide shows monthly sign-ups: January 100, February 120 and March 90. The teacher says, “The increase here did not continue into the following month,” while pointing to February. The recording then moves to another slide.
Select the evidence. Identify the requested time segment. Confirm that the chart appears during that interval instead of borrowing a similar slide from elsewhere.
Read the visible data. Extract the title, axis labels, units and values. If “sign-ups” is blurry, report the label as uncertain rather than silently changing it to revenue.
Connect narration and chart. Determine that “the increase here” refers to January-to-February, using timing and the visible pointer where available.
Calculate separately. A rise from 100 to 120 is 20%. A fall from 120 to 90 is 25%. The denominator changes, so these are not equal-and-opposite percentage changes.
Explain only what is supported. The figures show a rise followed by a fall. They do not establish why sign-ups fell.
Give a checkable answer. Include the relevant timestamp, the values read, the calculation and any missing evidence.
A good answer might say: “At approximately 02:18, the chart appears to show 100, 120 and 90 sign-ups for January, February and March. February is 20% above January; March is 25% below February. The explanation identifies a reversal but does not give a verified cause.”
A poor answer might say that March revenue fell because a marketing campaign ended. That invents both the quantity and the cause. More input channels do not justify more confident speculation.
Now remove the video and keep only the transcript. “Here” and “the following month” may become ambiguous. Remove the narration instead, and the model may still read the chart but lose the teacher’s intended emphasis. This comparison reveals which modality contributes to the task.
Where multimodal AI can be useful
Documents, diagrams and visual explanations
A student can ask about a graph beside a paragraph. A document-review system can connect a table entry to its footnote or flag a mismatch between a diagram and written instructions. These are possible applications, not guarantees of correct extraction. Small labels, complex layouts and ambiguous units still need checking.
Audio and video search
Cross-modal search can help locate a described event inside media. Google’s Gemini 1.5 technical report, introduced in 2024, studies long-context multimodal capabilities including video and audio tasks. Those reported experiments establish a research milestone, not a promise that every video question is answered correctly or that the same model versions remain available today. [6]
Accessibility and assistance
Describing an image, producing captions or explaining an unfamiliar interface can make information easier to access. The output should preserve uncertainty and give the user control over how much detail is provided. A general-purpose image description should not be treated as a dependable navigation or emergency-warning system.
Creative and editorial work
A creator might compare a storyboard with a brief, organise footage or check whether a caption refers to the correct image. Understanding media and generating media are different capabilities. A tool that explains an image does not necessarily edit it, and a tool that generates a video does not necessarily verify its factual content.
What can go wrong with multimodal AI?
Object hallucination: a system may describe something that is not in the image. The 2023 POPE paper studied this problem in large vision-language models and proposed a way to evaluate whether models incorrectly claim objects are present. Its findings concern the models and settings tested, not a universal failure rate for every model today. [7]
Lost detail: preprocessing can resize images, select frames or divide audio into chunks. A small handwritten digit or brief event may not survive the representation that reaches the model. Ask what the actual input processing preserves before concluding the model inspected every detail.
Conflicting evidence: a slide may show one number while narration gives another. The useful answer identifies the disagreement and points to both sources. Quietly choosing the more plausible version conceals an important limitation.
Overinterpretation: a pause, facial expression or tone does not reliably establish a person’s private intentions. Describing observable behaviour is different from asserting emotion, honesty or competence.
Privacy and untrusted content: recordings can include faces, background conversations and confidential screens. Use material you are authorised to process, minimise unnecessary personal information and check the service’s current data-handling terms. Text embedded in a screenshot is task data; it should not automatically become an instruction to send files or change accounts.
How to evaluate a multimodal system
A single overall score rarely answers whether a system suits your use case. MMMU is a research benchmark containing around 11,500 multimodal questions across 30 subjects, using varied image types such as charts and diagrams. It evaluates demanding perception and subject reasoning; it is not a complete test of audio, video, privacy or operational reliability. [8]
For the lesson example, create a small evaluation set with known answers and separate perception from calculation. If the model misreads 90 as 900, correct arithmetic on the wrong number is still a failed answer. If it reads the values correctly but uses the wrong denominator, the failure is elsewhere.
A practical evaluation checklist
Check
Useful test
Perception
Read labels and numbers at realistic recording quality
Alignment
Match a spoken reference to the right slide and timestamp
Reasoning
Verify calculations and distinguish observation from cause
Missing evidence
Include silent clips, blurry slides and incomplete diagrams
Contradictions
Make narration and a visible value disagree
Efficiency
Measure review time, latency and total processing cost
Also compare against simpler baselines: transcript-only, image-only and OCR plus a text model. If adding video changes an answer, check that the change comes from useful visual evidence rather than noise. If the extra modality adds cost without helping, it may not belong in that workflow.
Frequently asked questions
Is a PDF a separate modality?
A PDF is a file format. It may contain selectable text, scanned images, tables or diagrams. The relevant modalities depend on its contents and how the system processes them.
Does a multimodal model see like a person?
No. It processes learned representations. Human-style language about “seeing” is shorthand and should not imply identical perception or understanding.
Can it analyse every moment of a video?
That depends on the system’s processing method and limits. Ask about frame selection, audio handling and timestamp precision. A fluent summary alone does not establish complete coverage.
Is multimodal AI the same as agentic AI?
No. Multimodality concerns information types; agency concerns selecting and carrying out steps. A system can have either capability or combine them. See our agentic AI guide for the distinction.
What should a beginner try?
Use a non-sensitive image or diagram with known answers. Ask the system to separate visible facts, calculations and uncertain interpretation. Check the result against the original before trying a more complicated task.
The practical takeaway
Multimodal AI is useful when different kinds of evidence need to be connected. The essential questions are what information reaches the model, which relationships it can use, and whether its answer is checkable. More modalities expand possibilities; careful evaluation determines whether they improve the result.
You can read a hundred books and still feel unwise when life gets hard. Real wisdom is not information. It is the ability to see clearly, decide well, and stay steady when things go wrong.
These five practices help you develop that ability without complicated systems.
1. Pause Before You React
Most mistakes happen in the first three seconds. When something bothers you, count to five before speaking or acting. This small gap is where wisdom begins.
2. Ask “What Can I Actually Control?”
Write the problem. Then circle only the parts you can influence. Focus your energy there. Everything else is noise.
3. Review Your Day in Three Minutes
Each evening ask:
What went well?
What did I do that I would avoid next time?
What will I do differently tomorrow?
This simple review compounds faster than any course.
4. Learn From People Who Disagree With You
Wisdom grows when you can understand a view you do not hold. Once a week, read or listen to someone whose opinion challenges yours. You do not have to agree. You only have to understand.
5. Protect One Quiet Hour
Wisdom needs space. Guard one hour a day with no notifications. Use it for thinking, reading, or walking. That hour is where clear judgment is built.
Quick Summary
Wisdom is built through small daily habits, not grand moments. Pause, focus on what you control, review, stay open, and protect quiet time. Start with one practice today. The rest will follow.