DEV Community: Noah Taro

Cutting ASR Inference Cost with NVIDIA MPS on Amazon EC2

Noah Taro — Mon, 31 Aug 2026 19:38:06 +0000

When an ASR pipeline is pushed to production, the interesting question is not only how fast it runs, but how much throughput you can extract from each GPU before latency starts to break. In the setup described here, that tradeoff was the main lever for reducing inference cost by 75% using NVIDIA MPS on Amazon EC2.

This post is a collaboration between AWS, NVIDIA, and Heidi. It also includes input from Jerron Chua, a Deep Learning Architect at the Generative AI Innovation Center at Amazon Web Services (AWS).

Why GPU sharing matters for inference

If you are running speech recognition at scale, a single model instance per GPU is often not the most efficient use of hardware. There are a few common ways to share GPU capacity:

Time-slicing, where work from multiple processes is interleaved on the same device
MIG, which partitions supported GPUs into isolated slices
MPS, which allows multiple CUDA clients to run concurrently through a single GPU context

The key detail here is that NVIDIA CUDA MPS is a binary-compatible alternative implementation of the CUDA API. That matters because it lets existing CUDA-based workloads benefit from concurrent execution behavior without rewriting the application around a different programming model.

For inference workloads, the practical question is whether that concurrency improves utilization without pushing latency outside the acceptable range. That is where the rest of the setup comes in.

The inference pipeline on EC2

The pipeline in this work runs on Amazon EC2 g6e.4xlarge and g7e.4xlarge instances, both using NVIDIA L40S GPUs with 48 GB of memory. The architecture is built from three containerized components, which keeps the deployment modular and easier to reason about when measuring performance and tuning concurrency.

That container-based layout is important for a few reasons:

It isolates responsibilities across the pipeline.
It makes it simpler to swap configurations while keeping the inference flow comparable.
It helps with repeatable benchmarking when evaluating whether MPS is actually improving efficiency.

The article’s focus is not on a brand-new inference stack. It is on taking a practical deployment, then using MPS to improve the cost-to-throughput balance.

What you need before building the image

A key prerequisite is the NVIDIA Triton Inference Server container:

nvcr.io/nvidia/tritonserver:26.03-py3

That base image anchors the serving side of the workflow. From there, the setup builds an all-in-one image for the specific model package used in the test.

The build command is:

docker build -f Dockerfile.single \
  --build-arg LOCAL_NEMO_FILENAME=your_model.nemo \
  -t parakeet-mps:latest .

A few implementation details are worth calling out for builders:

The Dockerfile is named Dockerfile.single.
The model file is passed in as a build argument through LOCAL_NEMO_FILENAME.
The resulting image is tagged as parakeet-mps:latest.

This is a straightforward pattern if you already package models into containers for inference. The main thing is that the model artifact is baked into the image build flow, which makes the runtime environment more reproducible when you compare different GPU-sharing strategies.

Triton plus MPS on g6e.4xlarge

One of the evaluated configurations is Triton + MPS on g6e.4xlarge. The important takeaway from the benchmark is not just that MPS improves utilization, but where the operating point lands once latency is considered.

The source identifies the optimal operating point as the last concurrency level where mean latency stays below 650 ms and p99 stays below 1,000 ms.

That distinction matters because raw throughput alone can be misleading. A configuration can process more requests if you keep increasing concurrency, but if the tail latency climbs too far, it stops being useful for many production ASR scenarios. The benchmark therefore uses a practical latency boundary instead of treating every additional request in flight as a win.

For developers tuning a similar system, this gives a useful rule of thumb:

Increase concurrency until utilization improves.
Watch both mean and p99 latency.
Stop at the last point that still satisfies the service objective.

That approach is more operationally relevant than maximizing throughput in isolation.

Why the result is meaningful

The headline result is a 75% reduction in inference cost. The reason this is interesting is not that the number itself is magical, but that it comes from improving how the GPU is shared rather than replacing the model or redesigning the full pipeline.

That makes the technique attractive when:

The model is already working correctly.
The bottleneck is inefficient GPU usage rather than model quality.
You need better economics without changing the core ASR behavior.

At the same time, MPS is not a universal answer. The same latency constraints that make the configuration viable also limit how far you can push concurrency. If the workload becomes too crowded, latency moves outside the acceptable range and the benefit disappears. That is why the benchmarked operating point matters more than a generic promise of “more throughput.”

Operational cleanup matters too

There is one practical detail that is easy to overlook after benchmarking: clean up the resources.

In this setup, that includes deleting attached Amazon EBS volumes, which may contain:

model checkpoints
TensorRT cache

For anyone running iterative performance experiments on EC2, this is not just housekeeping. Leftover volumes can continue to incur cost after the test is over, and they can also leave behind artifacts that confuse future runs if you are trying to reproduce results cleanly.

Takeaway for builders

If you are tuning ASR inference on EC2, the useful lesson from this setup is that GPU sharing can be a cost lever when it is evaluated against explicit latency targets. NVIDIA MPS, used with Triton on L40S-backed EC2 instances, can improve the economics of inference, but only when you choose the concurrency level that still respects both mean and tail latency.

In other words, the optimization is not “run more at any cost.” It is “find the highest concurrency that still behaves like a production service.” That is what makes the 75% cost reduction operationally meaningful rather than just a benchmark number.

Stop Being a “Meat Proxy”: A Practical Workflow for Using AI Without Outsourcing Your Judgment

Noah Taro — Fri, 28 Aug 2026 11:22:05 +0000

A common mistake in AI-assisted work is assuming that speed alone equals productivity.

If a chatbot drafts something and you paste it into Slack, email, a spec, or a code review without reading it carefully, you are not saving time in any meaningful way. You are just moving the thinking step from your own brain to the recipient’s inbox. That is the behavior developer Niklas Gruhn called being a “meat proxy,” and the term lands because it describes a real failure mode: a human acting as a thin wrapper around machine output.

The problem is not AI itself. The problem is unedited AI output being treated as finished work.

For developers, this shows up everywhere:

replying to a teammate with a raw chatbot answer instead of a considered response
copying generated explanations into docs without checking accuracy
pasting AI-written code or configs without understanding the assumptions
delegating too much context to the model and too little to yourself

That pattern is closely related to what researchers call cognitive offloading or cognitive surrender. In practice, it means your judgment, not the model’s text, is the scarce resource you are failing to apply.

Why this matters in developer workflows

Most software work is contextual. A suggestion can be syntactically correct and still be the wrong answer for your codebase, your team norms, your performance constraints, or your threat model.

AI tools are very good at producing plausible text. They are much less reliable at knowing:

your internal architecture
which edge cases are important in your product
whether a dependency is acceptable
what your team considers readable or maintainable
which tradeoffs matter more right now

That is why blindly forwarding model output often creates more work downstream. The recipient has to re-evaluate the response from scratch, and now they also have to guess whether you understood it.

From a collaboration perspective, that is the real issue. Not only is it low value, it can reduce trust. If people know they still need to verify everything you send, you have not accelerated the workflow. You have added another review layer.

A better pattern: use AI as a draft generator, not a context replacement

Gruhn’s practical advice is simple and still the best starting point: read it, understand it, validate it, then rewrite it in your own words.

That sequence is worth turning into a repeatable workflow.

1. Ask AI for a draft, not the final artifact

Use the model to generate options, summaries, or a first pass. This is especially helpful for:

outlining a response
surfacing missing considerations
translating between formats
brainstorming edge cases
generating a rough explanation for a nontechnical audience

The key is to treat the output as raw material.

2. Validate every claim that matters

Before you send anything, check the parts that could affect correctness:

facts
API behavior
security implications
compatibility assumptions
performance claims
version-specific details

If the output contains code, do not rely on confidence in the prose. Read the code line by line. Ask whether it handles error paths, types, boundaries, and dependency constraints. A polished explanation does not make an untested snippet safe.

3. Re-express the answer in your own voice

This is the step people skip most often, and it is also the step that adds the most value.

Rewriting is not cosmetic. It forces you to decide:

what actually matters
what can be omitted
what should be emphasized
what tone is appropriate for the audience

That is where your judgment becomes visible. A teammate does not need a transcript of a chatbot conversation. They need your interpretation of the issue.

4. Add context the model cannot infer

If you want to be useful, include the thing the model cannot know on its own:

“We cannot introduce a new dependency here.”
“This needs to fit our existing event pipeline.”
“Keep the explanation short because this is for the support team.”
“We already tried approach A and it failed because of X.”

That small amount of context often makes the difference between generic slop and genuinely useful work.

A simple standard you can apply today

If you want a rule that is easy to remember, use this:

Never send model output you would not be willing to defend in a meeting.

That standard is strict enough to catch most meat-proxy behavior.

Before sending anything AI-assisted, ask yourself:

Could I explain this without looking at the tool output?
Do I understand why this answer is correct?
Did I check the parts most likely to be wrong?
Did I tailor this to the actual recipient?
Would this still be useful if the model were unavailable?

If the answer to any of those is no, you probably need another pass.

Tradeoffs: where AI helps and where it hurts

There is a real upside to AI in engineering workflows. It can speed up first drafts, reduce blank-page friction, and help you explore alternatives faster than manual searching alone.

But the tradeoff is obvious: the easier it becomes to produce text, the easier it becomes to stop thinking.

That risk is especially visible in developer culture because so much of our work is already mediated by tools. We rely on autocomplete, linters, formatters, search, and codegen. None of those are inherently bad. The difference is whether the tool amplifies your understanding or replaces it.

Autocomplete assists the human. A meat proxy workflow inverts that relationship.

The practical takeaway

The useful habit is not “never use AI.” It is “do not let AI stand between you and your own judgment.”

Use it to draft, compress, and explore. Do not use it to avoid reading, thinking, or contextualizing. If you are sending something to another person, make sure the final message reflects your understanding, not just the model’s phrasing.

That is the difference between using AI as a tool and becoming the tool.

Use Four Result States for Safer Publishing Workflows

Noah Taro — Tue, 25 Aug 2026 09:36:44 +0000

A boolean success flag is not expressive enough for real publishing automation. A submission can be confirmed, explicitly rejected, blocked before submission, or submitted without enough evidence to determine the final result.

Those cases should map to four states:

confirmed success,
confirmed failure,
blocked before side effect,
pending confirmation.

The last state is especially important. Automatically republishing an unresolved submission can create duplicates. Instead, the verifier should visit the account page, look for a recent item with the expected author and content, and only then decide whether a retry is safe.

This state model makes records more accurate and makes recovery behavior much easier to reason about.

How to Prevent Exact-Count Automation From Overshooting

Noah Taro — Tue, 25 Aug 2026 09:36:14 +0000

Suppose a task requests five reactions. The automation clicks the first item, but the page does not expose a confirmation signal quickly enough. If the workflow immediately moves to the next item, the first action may still succeed later. The final report can show four confirmed reactions while the account actually made five or six.

The fix is a side-effect-aware retry policy:

Retry when no action was taken.
Count only confirmed state changes.
Stop after an action whose result is unresolved.
Resume only after independent verification resolves the uncertainty.

This policy sacrifices a little apparent completion rate in exchange for exactness. In production operations, that trade is usually correct: a visible shortage with a precise reason is safer than a hidden overshoot.

Build Automation Around Observable State, Not UI Clicks

Noah Taro — Tue, 25 Aug 2026 09:35:18 +0000

UI automation becomes fragile when a click is treated as the outcome. The browser may dispatch the event while the application rejects the request, delays the update, or replaces the original element.

A more reliable workflow has four stages:

identify a safe target using stable structure,
perform one visible action,
wait for the application to settle,
verify a durable state change.

For a reaction, that durable state might be a selected control or an increased count. For publishing, it should be the new public item with the expected content. The verification result, not the click call, should drive the final status.

This design also produces better diagnostics. Instead of a generic failure, the system can say whether it never found a target, could not perform the action, or performed it but could not confirm the outcome.

Choosing a Sentiment Analysis Tool in 2026: A Workflow-First Guide for Builders

Noah Taro — Mon, 24 Aug 2026 20:03:45 +0000

When a team starts evaluating sentiment analysis software, the first question is usually not “which vendor has the fanciest AI?” It is “how will this actually fit into the way we work?”

That distinction matters. Sentiment analysis is only useful when it helps a team do something concrete: catch a negative trend early, separate signal from noise, route conversations to the right owner, or turn raw mentions into something decision-ready. If the tool cannot support one of those workflows, the dashboard may look impressive without changing outcomes.

This guide takes a practical angle: what sentiment analysis tools do, how the pipeline works, where the main tradeoffs show up, and what to test before you commit.

What sentiment analysis tools are good at

At the core, a sentiment analysis tool uses AI and natural language processing to classify mentions as positive, negative, or neutral. The inputs can come from social platforms, forums, news, reviews, blogs, and other public web sources.

The value is not just in labeling tone. Better tools also detect:

sarcasm and irony,
emotion such as joy, anger, fear, or surprise,
intent like complaint, question, recommendation, or purchase interest,
and aspect-level sentiment, such as shipping versus product quality.

That matters because a single post can be positive about your product and negative about your delivery experience. If your platform collapses everything into one score, you lose the nuance that support, product, and communications teams actually need.

How the analysis workflow usually works

Most tools follow the same basic sequence:

1. Collect mentions

The platform gathers brand, product, or keyword mentions from the sources it supports.

2. Analyze language

Models inspect wording, context, emojis, phrasing, and sometimes the conversation around the mention.

3. Classify sentiment

Each mention is tagged positive, negative, or neutral. More advanced systems also assign intensity scores or emotion labels.

4. Turn output into action

The results are displayed in dashboards, alerts, reports, or workflow queues so teams can act on them.

From a builder’s perspective, this is where implementation choices start to matter. Some tools are designed for continuous monitoring and alerting. Others are better for ad hoc analysis. A few are meant to live inside a larger publishing, engagement, and reporting workflow.

That difference affects everything from how quickly you can respond to an issue to whether the data is useful outside the marketing team.

The main types of sentiment analysis

If you are shortlisting tools, it helps to know which type of analysis you actually need.

Fine-grained sentiment

Instead of only “positive” or “negative,” this gives you a scale. It is useful when you care about smaller shifts over time.

Aspect-based sentiment

This splits one mention into separate attributes such as price, support, quality, or shipping. Product and customer experience teams usually get the most value here.

Emotion detection

This goes beyond polarity and identifies feelings such as anger, fear, joy, or surprise. It is especially useful during launches and crises, when a shift from surprise to anger tells you more than a small drop in score.

Intent-based sentiment

This looks at what someone wants next, such as a complaint, a question, a recommendation, or a buying signal. That is often what support and sales teams really need.

The practical takeaway is simple: if a tool only gives you polarity, make sure that is enough for your use case before you buy.

What to evaluate before you choose

A lot of buyers compare pricing first and accuracy later. That often goes wrong. A cheaper tool that misses sarcasm, ignores the wrong sources, or cannot feed data into your workflow can cost more in missed decisions than a premium platform.

Here is the checklist I would use in a real rollout.

Test accuracy on your own data

Do not rely on the sales demo alone. Test the platform against your own mentions, including slang, negation, sarcasm, and short replies. This is where basic models fail first.

Verify source coverage

Check whether the tool covers the exact places your audience uses:

major social platforms,
review sites,
forums,
blogs,
online news,
and the languages you need.

If your customers spend time in one niche community and the tool cannot ingest it, the dashboard will look polished but incomplete.

Decide between real-time and periodic analysis

If you need to respond to a growing issue, you want alerts and live monitoring. If you only need a monthly brand report, batch analysis may be enough.

Look at integration depth

This is the part builders often care about most. Ask whether sentiment data can flow into:

a CRM,
support tooling,
publishing workflows,
BI dashboards,
or reporting systems.

If the answer is only “you can export CSVs,” the workflow is probably not fully connected.

Check governance and access control

For enterprise teams, the important question is not only “can it analyze sentiment?” but also “can we safely use the output?”

Look for:

role-based access,
audit trails,
approval workflows,
data residency options,
and clear handling of customer data.

Understand the pricing model

Sentiment tools are priced very differently:

per user,
per keyword,
per mention,
per account,
or via custom enterprise quote.

That matters because usage can spike quickly during launches or incidents.

Standalone listening tools versus all-in-one workflows

One of the biggest decisions is whether you want a standalone listening tool or sentiment built into a broader social system.

Standalone products often win on depth. They may cover more sources, support richer analysis, and provide stronger listening features.

All-in-one platforms often win on workflow. If your team publishes, responds, monitors, and reports from the same place, you reduce context switching and make it easier to act on insights.

In Hootsuite Social OS, sentiment analysis, listening, publishing, and reporting are connected through Lumen and Perch. The benefit is not just having a chart. It is being able to spot a shift, understand it, and hand it off inside one system.

That workflow matters most when multiple teams need to work from the same signal.

Where LLMs fit, and where they do not

Yes, tools like ChatGPT can classify sentiment on text you paste in. They are useful for quick reads or one-off interpretation.

But there is a big difference between ad hoc analysis and operational monitoring.

General-purpose LLMs are useful when:

you already have the text,
you want a quick interpretation,
and you do not need historical tracking or alerts.

Dedicated sentiment tools are better when you need:

continuous collection,
consistent scoring,
trend tracking over time,
alerts,
and a governed workflow.

That last point is often the deciding factor. Pasting customer conversations into a consumer chatbot is usually not something legal or security teams will approve.

A practical way to narrow the field

If I were choosing a tool today, I would frame it like this:

Need enterprise-wide visibility? Prioritize source coverage, governance, and integration.
Need product or CX insights? Look for aspect-based analysis and strong context handling.
Need a lightweight read for a small team? Simpler tools can be enough if you accept limited depth.
Need creator or campaign reporting? Look for tools that make it easy to compare time periods, hashtags, and mentions.
Need one-off checks only? Free analyzers may be enough for a quick snapshot.

The core tradeoff is always depth versus workflow fit. A more advanced platform is not automatically better if nobody can operationalize the output.

Common failure modes to test

Before you sign a contract, stress test these edge cases:

sarcasm and irony,
negation like “not bad,”
mixed sentiment in a single post,
multilingual language handling,
and region-specific slang.

If the platform struggles with these, make sure that limitation is acceptable for your data set.

Final thought

The best sentiment analysis tool is the one that matches your workflow, not the one with the longest feature list.

If you are choosing for a small team, a simple system with clear charts may be enough. If you are choosing for an enterprise environment, the real value is often in how well sentiment data moves through the rest of your stack, from monitoring to reporting to response.

That is the decision worth making carefully, because sentiment only matters when it changes what your team does next.

How to Grow Instagram Followers in 2026 by Fixing the Real Bottleneck

Noah Taro — Sat, 22 Aug 2026 01:11:33 +0000

Most Instagram growth advice begins with the same assumption: if follower growth is slow, you just need more Reels, better hashtags, or stricter consistency.

That assumption is incomplete. An account can get views without creating interest, profile visits without converting them into follows, or followers that never come back. In other words, the problem is rarely “not enough posting” by itself. It is usually a broken stage in the growth chain.

A more useful way to think about Instagram growth in 2026 is as a sequence:

Discovery → Interest → Follow conversion → Retention

When growth stalls, one of those stages is usually the bottleneck. If you fix the wrong one, you can work harder and still see little progress.

Diagnose the Bottleneck Before Changing Tactics

Before editing your content calendar, figure out where people are dropping off.

A simple diagnostic view looks like this:

What you see	Likely bottleneck
Low reach from non-followers	Discovery is weak
Reach is decent, but profile visits are low	The post is not creating enough interest
Profile visits are high, but follows are low	The profile is not convincing enough
Followers increase, but engagement drops later	New followers are not a good fit
Followers rise, but interaction stays flat	The content may be attracting the wrong audience

This is not an official Instagram formula. It is a debugging framework that helps you stop guessing.

For example, if your content is not being recommended to non-followers, changing posting time will not solve that problem. If your profile gets visits but nobody follows, more content volume will not fix weak positioning. The next step only makes sense after you know which stage is failing.

Instagram also has platform-level recommendation logic worth checking. Public content may be eligible to appear in Reels, Explore, Search, Suggested Accounts, and Feed recommendations, and professional accounts can review recommendation status through Account Status. If a post, profile photo, or bio conflicts with Instagram’s Recommendations Guidelines, that can suppress reach before your tactics even matter.

Discovery: Reach People Who Actually Care

A lot of accounts chase broad attention because broad attention feels like progress. Views go up, but growth does not.

The real goal is not to reach everyone. It is to reach people who are likely to want more of your content later.

Build topics around repeated audience problems

The best content usually comes from questions, pain points, and misunderstandings your audience already has.

A web designer could post about running a business in general, but that is too vague to drive follow intent. A post about three homepage mistakes that reduce inquiries is more specific, more useful, and easier to connect to the account’s core expertise.

A practical starting point is to define three to five content pillars around recurring needs. For a web design account, those might include:

Website conversion problems
Before-and-after redesigns
Mobile usability and basic SEO
Client project breakdowns
Design decisions behind the final result

Specificity matters, but only when it still leaves room for future posts. The goal is not to be narrow for its own sake. The goal is to become clearly useful inside a stable area of expertise.

Match format to the growth goal

Different formats do different jobs:

Format	Best used for	Signals to review
Reels	Introducing an idea to unfamiliar viewers	Non-follower reach, watch time, shares, profile visits
Carousels	Explaining steps, comparisons, or detailed lessons	Saves, shares, profile visits
Static posts	Showing products, portfolios, or case studies	Saves, relevant comments, profile visits
Collab posts	Reaching the audience of a related creator, brand, or customer	Non-follower reach, profile visits, follows
Stories	Maintaining relationships with current followers	Replies, link taps, poll responses, repeat views

Reels can support discovery, but they are not a universal answer to Instagram follower growth. Educational, visual, and portfolio-based accounts often get stronger profile activity from carousels or image-led posts.

The format should be chosen by the action you want, not by whatever is currently fashionable.

Make discoverability easier with Instagram SEO

Instagram SEO is mostly about helping the platform and the viewer understand what your post covers.

Use language your audience would actually search for or recognize:

Put a role, service, or topic in the Name field
State the subject naturally in the caption
Add a clear title on the cover or first frame
Use on-screen text that explains the topic without sound
Include location terms if you serve a local market
Use a small set of closely related hashtags

“Three website mistakes that cost small businesses leads” is stronger than “You need to see this.” The first version tells people who it is for, what it covers, and why it matters before they even tap.

Hashtags can still help with topic clarity, but relevance matters more than quantity. Broad tags like #viral or #explorepage usually add noise rather than useful context.

Package ideas so they are worth saving and sharing

Once the topic is clear, make the value obvious.

Instead of “Tips for a better Instagram bio,” try “A four-line Instagram bio structure for local service businesses.” The second version gives the reader a concrete outcome and a reason to keep the post.

Content that tends to get saved or shared often looks like:

A short process or checklist
A mistake-and-fix breakdown
A comparison or before-and-after example
A reusable template, framework, or series

A direct request to save or share does not rescue a weak post. The post itself has to deserve that action.

Turn Profile Visits Into Follows

A post can perform well and still fail to grow the account. For a visitor to follow, they need to see a clear reason to stay.

That means each post should connect to a broader promise from the account.

A fitness coach can turn one exercise tutorial into a mobility series. A marketing account can connect one ad example to a pattern of campaign breakdowns. A designer can use a finished project to lead into the process behind the result.

A generic “follow for more” CTA is easy to ignore because it does not define what more actually means. More effective CTAs name the future value:

Follow for weekly landing page breakdowns
See the pinned guide for the full process
The next post covers the mistakes to avoid
More examples are saved in the Case Studies highlight

Make the profile answer the obvious questions fast

A strong profile should answer four questions in seconds:

Who is this for?
What value will followers receive?
What is the main topic?
What should the visitor do next?

A vague bio like “Helping you become your best self” does not do that. A clearer version works better:

Strength training for busy professionals. Weekly 20-minute workouts and practical form tips.

A good Instagram bio for more followers does not need clever language. It needs a clear audience, subject, and value proposition.

The recent grid should reinforce the same promise. If the bio says one thing and the feed shows something else, conversion drops.

Pinned posts and Highlights can support the decision to follow as well. A simple structure is:

Start Here: who the account serves and what it covers
Best Example: a strong tutorial, case study, or series
Next Step: proof, a resource, or a useful follow-up

Highlights can organize recurring value under labels such as Tips, Results, FAQs, Services, or Behind the Scenes.

Keep New Followers Interested

Growth is not only about getting the follow. It is also about keeping the right people engaged after they arrive.

The easiest way to lose new followers is to abandon the content that earned the follow in the first place.

If a tutorial performed well, build on it with a deeper guide, a case study, or a mistake-and-fix post. If a comparison post worked, turn it into a series for different use cases or budgets.

This also applies to promotional content. It should still support the account’s main purpose. An agency can present a service through a campaign breakdown instead of replacing educational content with repeated sales posts.

Stories, comments, and DMs are useful because they reveal what your audience keeps asking for. Repeated questions, objections, and requests for examples are direct signals for the next post.

Unfollows can also teach you something, but only when you look for patterns. A sharp increase may follow a major topic shift, too many promotional posts, a sudden change in publishing frequency, or viral content that attracted the wrong audience.

The goal is not to keep every follower. The goal is to keep the people the account is actually built for.

Use Insights to Test One Change at a Time

Instagram Insights become useful when they help you answer one specific question.

Bad test: “Let’s improve engagement.”

Better tests:

Does this opening improve watch time?
Does this topic increase profile visits?
Does this CTA produce more follows?
Does this format generate more saves or shares?

The cleaner the test, the easier it is to learn from it. If the topic, format, cover, caption, timing, and CTA all change at once, you will not know what caused the result.

A better approach is to keep most of the structure stable and adjust one major variable. For example, test two different opening hooks on the same Reel topic, or compare two covers for similar carousels.

When reading results, follow the chain:

Non-follower reach → Profile visits → Follows → Return engagement

That sequence usually shows where the slowdown happens.

If reach grows but profile visits do not, the post got attention without creating enough account interest. If profile visits rise but follows do not, the profile positioning is likely weak. If follows rise but return engagement stays flat, the account may be attracting low-fit followers.

Scale Only After the Process Works

If you manage multiple Instagram accounts, it is tempting to scale before the workflow is stable. That usually multiplies noise instead of growth.

Scale only after one account has a repeatable process for:

Defining the audience
Choosing content pillars
Matching format to goal
Converting visits into follows
Testing changes consistently
Reviewing results in a structured way

Once that process works, a team can adapt it for different brands, clients, or regions without making every account look identical.

For authorized multi-account operations, DICloak can keep each account in a separate Browser Profile with its own cookies, login session, and browser data, while also supporting role-based access, operation logs, grouping by client or region, and batch profile management. That is useful after the underlying content system is already working.

Final Takeaway

If Instagram growth feels stuck, do not start with more hashtags or more posting frequency. Start by asking which stage is failing: discovery, interest, follow conversion, or retention.

Once you know the bottleneck, the next move becomes much clearer. That is usually the fastest way to turn views into followers, and followers into an audience that keeps coming back.

Instagram Hashtags in 2026: The Mistake Most Teams Make, and the Workflow That Fixes It

Noah Taro — Wed, 19 Aug 2026 12:37:22 +0000

If your Instagram posts are not getting traction, the most common assumption is that you need more hashtags.

That is usually the wrong starting point.

In 2026, the bigger problem is not shortage, but mismatch: too many tags, too many generic words, and too little discipline in how they are chosen. Instagram has become much better at understanding images, captions, and user intent. That means hashtags still matter, but they work best as a precise signal, not as a volume game.

For builders, marketers, and operators, the practical question is simple: how do you turn hashtags into a repeatable workflow instead of a guess?

The common mistake: treating hashtags like a slot machine

A lot of accounts still behave as if more tags automatically means more reach. That idea made sense years ago, but it creates noise now.

The first failure mode is over-tagging. If you dump 30 unrelated hashtags into a post, the system gets a muddled signal about what the content is actually about. A post about a modern desk lamp should not be surrounded by broad or random tags like #love, #puppy, or #happy. Those tags do not describe the content, and they attract the wrong audience.

The second failure mode is chasing popularity instead of relevance. A tag like #SummerVibes on a winter coat post might get clicks, but from the wrong people. They are searching for beaches and cold drinks, not outerwear. That mismatch hurts engagement, and low engagement is a signal that your content is not useful for that audience.

The third failure mode is ignoring restricted or banned hashtags. Instagram does hide certain tags because they have been abused by spam or low-quality content. Some of those words are not obviously bad at first glance, which is why checking matters before publishing.

A practical rule for 2026: use 3 to 5 highly specific hashtags

The cleanest operational rule is to keep each post to 3 to 5 hashtags.

That is not about playing it safe for its own sake. It is about forcing precision. When you only have a few slots, you cannot waste them on vague terms. You are pushed to select the hashtags that best describe the image, the product, and the audience intent.

For example, if you sell custom neon signs for bedrooms, broad tags like #lights are too generic to be useful. More specific tags such as #CustomNeonSigns, #PinkBedroomDecor, or #TeenRoomIdeas are much better because they align with what the content actually shows and what a buyer may be searching for.

This smaller set also makes your captions look cleaner. That matters more than many teams admit. A caption packed with tags can read like spam. A short, focused set looks intentional and professional.

Build a hashtag workflow, not a one-off list

If you manage a brand account, the real goal is not finding one best hashtag set. The goal is building a repeatable selection process.

Here is a workflow that fits how Instagram works now:

1. Start from the post, not from a hashtag list

Ask what the post is actually about.

What is shown in the image?
What problem or interest does it relate to?
What would a potential buyer type into search?

This keeps the tags tied to content instead of trend-chasing.

2. Prefer niche tags over giant umbrella tags

Broad hashtags such as #fashion, #travel, or #food are so saturated that a post can disappear almost immediately. Niche tags have smaller audiences, but those audiences are more relevant and usually more engaged.

A tag like #VintageLeatherJackets is a better fit than #fashion if that is the item in the post. The audience is smaller, but the intent is much stronger.

3. Check whether a tag is safe before using it

Before you add a tag to your routine, search it inside Instagram. If the platform shows warnings or signs of restriction, do not use it.

This is a small step, but it prevents a common failure: publishing with a hidden-reach tag and then wondering why the post performs like it never existed.

4. Review performance by post, not by assumption

Instagram Insights matters because it shows whether the tags actually contributed to reach. The key metric to watch is reach from hashtags. Profile visits are also useful because they show whether those viewers were curious enough to click through.

If a post gets little or no hashtag reach, the problem may be tag relevance, not content quality.

How to test hashtag sets without guessing

Most teams do not need fancy experimentation. They need basic discipline.

A simple A/B testing routine is enough:

Post one piece of similar content with 3 hashtags.
Post another similar piece at a different time with 5 hashtags.
Keep the topic, style, and posting window as close as possible.
Compare reach from hashtags and profile visits after a week.

You are not trying to prove that one universal number is magic. You are trying to learn what your account responds to in practice. That is a much better use of time than copying someone else’s list.

Use tools as input, not as final decisions

There are useful tools for finding hashtags, but they should support your judgment, not replace it.

Instagram search itself is a solid starting point. When you type a keyword, it suggests related tags and shows how active they are.

Display Purposes is another useful free option because it can surface related hashtags and filter out banned ones automatically. That can save time during research.

Hashtag generators can also help, but only if you treat them as idea sources. Do not copy the whole output. Generate a long list, then manually filter it down to the few tags that truly match the post. If you type coffee, for example, a generator may return dozens of options, but a cold brew photo should not inherit hot coffee or winter-themed tags just because they appeared in the list.

If you pay for a tool, the features that matter most are:

daily search volume
a way to check restricted or shadowbanned tags
related secondary keywords

Those features help you choose with evidence instead of intuition.

What to avoid if you want stable reach

A few habits reliably hurt performance:

Using very broad, overused tags like #love or #photooftheday
Mixing in irrelevant trending tags just to borrow attention
Reusing the same tag list on every post
Ignoring monthly changes in search behavior
Using suspicious or restricted tags without checking them first

One more point: spammy engagement bait like #like4like or #followme attracts low-quality interactions, not customers. That may inflate activity in the short term, but it does not build a useful audience.

Seasonal and event hashtags still work, if they are specific

Seasonal tags can be effective because they match what people are already searching for during a given moment.

Examples include autumn tags like #FallDecorIdeas, spring tags like #SpringCleaningTips, or shopping-event tags like #BlackFridayDeals2026. These work because the user intent is immediate.

For events, you can also create your own branded tag. The key is to keep it short, readable, and easy to spell. A tag like #SummerGlowGiveaway26 is much easier to use than a long, awkward phrase. Before printing it anywhere, search it first so you know it is not already associated with something unrelated.

A simple operating model for the team

If multiple people touch your Instagram content, create a shared hashtag process:

maintain a shortlist of safe, relevant tags by content category
review that list monthly
log which posts received the best reach from hashtags
replace underperforming tags instead of keeping them forever

That is the real improvement most accounts need. Not more hashtags. Better process.

Final takeaway

The strongest Instagram hashtag strategy in 2026 is not about maximizing count. It is about precision, relevance, and review.

Use 3 to 5 tags. Make them niche-specific. Check for restricted tags. Measure reach from hashtags in Insights. Then update your list based on what the account actually proves, not what an old playbook recommends.

That workflow is simple enough to run every month, and disciplined enough to improve reach without turning your captions into noise.