Listen to this Post
A New Kind of Research Experiment
AI agents are moving beyond writing code, fixing bugs, and generating documentation. The more interesting shift is happening when researchers give an agent access to real computational resources and ask it to improve the research process itself.
That is the central idea behind this experiment: an AI agent was given access to a GPU cluster and a substantial computational budget to search for a better data mixture for training a 2-billion-parameter vision-language model (VLM).
The result is striking. With a budget of 11,000 H100 GPU hours, the agent searched for a better training-data mixture and found one that increased the aggregate benchmark score from 58.8% to 62.8%. That represents a 6.8% relative improvement over the human-designed baseline, while also improving performance across all ten evaluated benchmarks.
But the experiment is about more than squeezing another few percentage points out of a model.
It points toward a broader change in how machine-learning research may be conducted: AI agents can increasingly participate in the experimental loop itself, rather than simply helping researchers write the code around it.
The Real Problem Is Not Always the Model
Modern AI research often focuses on model architecture, parameter counts, training techniques, and compute. Yet the data used to train a model can be just as important.
A training dataset is rarely a perfectly balanced collection. Researchers must decide which sources to include, how much weight each source should receive, which examples deserve greater representation, and how those choices interact with the model’s eventual capabilities.
These decisions have traditionally depended heavily on human experience.
Researchers inspect datasets, run experiments, interpret benchmark results, form hypotheses, adjust the mixture, and repeat the process. It is an expensive and time-consuming feedback loop.
The autoresearch approach changes one part of that equation.
Instead of asking an agent merely to execute a predefined experiment, researchers can give it an objective and allow it to search through possible configurations.
Giving an Agent 11,000 H100 Hours
The headline experiment gave an agent a computational budget of approximately 11,000 H100 GPU hours.
Its task was straightforward to describe but considerably harder to execute: find a better data mixture for training a 2B VLM.
The baseline data mixture had been designed by humans. The agent then explored alternatives, using benchmark performance as feedback.
This creates a miniature automated research laboratory.
The agent proposes a change, runs the experiment, observes the result, compares it with previous attempts, and decides what to try next.
Rather than having a researcher manually perform every iteration, the agent becomes part of the optimization process.
The Result: 58.8% to 62.8%
The strongest data mixture discovered during the search increased the aggregate score from 58.8% to 62.8%.
That four-percentage-point absolute improvement is important because benchmark gains become increasingly difficult to obtain as systems improve.
More importantly, the reported improvement was not concentrated in a single benchmark.
The optimized mixture improved performance across all ten benchmarks used in the evaluation.
That makes the result considerably more interesting than a scenario where an optimization happens to exploit one particular test.
Why the Relative Improvement Matters
The experiment describes the improvement as approximately 6.8% relative to the human-built baseline.
That distinction is useful.
An increase from 58.8% to 62.8% is a 4.0 percentage-point absolute gain, but when compared proportionally with the original score, it corresponds to roughly 6.8%.
For model-development work, both numbers matter.
The absolute gain tells us how much the benchmark score moved. The relative gain provides another way of understanding the magnitude of the improvement compared with the original performance.
Data Mixture Optimization Could Become a Major Research Tool
The deeper implication is that data selection itself can become an optimization problem.
Instead of researchers deciding once how much data from each source should be used, an automated system can potentially search through thousands of combinations.
A data mixture might include different proportions of instructional examples, image-text pairs, synthetic data, web-derived material, captions, mathematical content, code, or domain-specific datasets.
The search space can become enormous.
Humans can propose sensible starting points, but agents can potentially explore combinations that would be tedious for a person to test manually.
The Agent Is Not Replacing the Researcher
It would be easy to interpret the experiment as evidence that researchers are becoming unnecessary.
That would be the wrong conclusion.
The agent still needs a well-defined objective, infrastructure, evaluation criteria, constraints, and human oversight.
Someone must decide what counts as success.
Someone must decide which benchmarks matter.
Someone must determine whether an apparent improvement represents genuine progress or simply optimization against a flawed evaluation setup.
The agent can search the space, but humans still need to define the space and judge the significance of what is found.
Research Becomes an Iterative Conversation With Machines
The more interesting model is therefore not human versus AI.
It is human plus AI in a tighter experimental loop.
A researcher can specify the goal, provide the environment, and establish the constraints. The agent can then generate hypotheses, implement experiments, run them, analyze outcomes, and propose subsequent experiments.
This creates a feedback loop that can operate much faster than traditional manual experimentation.
The result is a new form of research workflow in which the agent becomes something closer to an experimental collaborator.
The Same Principle Applies to Technical Writing
The article also makes a second argument that is easy to overlook.
The same agents that can modify code and conduct experiments can also dramatically change how technical research is communicated.
Writing code is increasingly cheap.
An agent can rewrite a paragraph, generate another explanation, produce an alternative visualization, rebuild a chart, or create an interactive diagram in seconds.
That means the scarce resource is shifting.
The difficult part is increasingly editorial judgment.
Attention Is the First Constraint
Readers do not have unlimited attention.
A technical article can contain excellent research and still fail if the reader has to navigate through several paragraphs of background before discovering why the result matters.
The most important result should therefore appear early.
In this experiment, the central statement is compelling enough to lead the article:
An agent received an 11,000-H100-hour budget to search for the best data mixture for training a 2B VLM.
That sentence immediately establishes the experiment, the scale, and the objective.
The reader then has a reason to continue.
Results Should Earn the
Once the reader understands what happened, the technical details become more meaningful.
The improvement from 58.8% to 62.8% provides the payoff.
Only afterward does it make sense to explain the architecture of the search, the experimental process, the dataset composition, and the computational infrastructure.
This is not merely a writing trick.
It is a way of respecting the
Visuals Often Determine What Gets Read
Technical readers frequently do not read an article from beginning to end.
They scan.
They look at charts.
They examine tables.
They inspect diagrams.
Then they decide which sections deserve deeper attention.
That makes visualization part of the argument rather than decoration added after the research is finished.
A well-designed figure can communicate the structure of an experiment before the reader has absorbed every paragraph describing it.
An Accurate Chart Can Still Tell the Wrong Story
Accuracy alone does not guarantee effective communication.
Imagine an experiment in which progress happens through several large jumps separated by long periods of little movement.
A conventional line chart might make the overall trajectory appear almost flat.
An ordered experiment plot, however, could make those jumps immediately visible.
Both representations may be mathematically correct.
But they do not communicate the same thing.
This is why visualization should be treated as an analytical decision.
AI Makes Visualization Iteration Cheap
This is one area where coding agents are particularly useful.
A researcher no longer needs to spend an afternoon manually rebuilding a chart to answer a simple question such as:
Would this result be clearer as a timeline, a scatter plot, a progression diagram, or a comparison table?
An agent can generate several alternatives.
The human researcher can inspect them and select the version that communicates the evidence most effectively.
That changes the economics of experimentation with presentation itself.
From Static PDFs to Interactive Research
Traditional papers have a major limitation: the final document is essentially a frozen representation of the research.
A PDF cannot dynamically respond to the reader.
A web-based technical article can.
A reader might hover over a benchmark and see its description.
A visualization can switch between comparisons.
A diagram can expose additional information.
An interactive chart can allow the reader to examine individual experiments.
The article argues that technical communication should increasingly take advantage of these capabilities.
Interactive Research Is More Than Visual Polish
Interactive presentation can also improve scientific understanding.
Suppose a benchmark table contains ten evaluations.
In a PDF, the reader may need to locate another section to understand what each benchmark measures.
On the web, that information could appear directly beside the relevant result.
The distance between evidence and explanation becomes smaller.
That matters because every additional cognitive step creates another opportunity for misunderstanding.
The Hugging Face Research Template
The article points toward the Research Template available through Hugging Face Spaces as an example of this philosophy.
The concept combines technical writing with web-native elements such as interactive diagrams, mathematics, citations, and PDF export.
It is influenced by the interactive research-writing tradition popularized by Distill.
The underlying philosophy is simple:
Research published on the web does not have to behave like a PDF displayed inside a browser.
Technical Writing Needs Its Own Quality Control
There is also a warning hidden inside the enthusiasm about AI-generated research and writing.
Fluent language can conceal mistakes.
An agent might move from discussing an eight-GPU machine to referring to a “GPU slot” without explaining the transition.
A paragraph can sound professional while containing an undefined term.
A chart can look beautiful while contradicting a number elsewhere in the article.
A benchmark can be mentioned without adequately explaining what it measures.
AI makes these mistakes easier to produce at scale.
AI-Generated Prose Requires AI-Level Skepticism
The appropriate response is not to reject agent-generated writing.
It is to review it with the same discipline used for agent-generated code.
Every logical step should be traceable.
Every important term should be defined.
Every number should reconcile with the underlying experiment.
Every claim should have appropriate evidence.
Every baseline should be identified.
Every figure should be introduced in the prose.
The smoother the writing becomes, the more important this verification process becomes.
Citation Discipline Still Matters
A technical blog does not necessarily need the citation density of an academic paper.
But that does not mean citations are optional.
Previous work should be acknowledged.
External announcements should be linked.
Baselines should be named.
Claims should be attributable.
Readers should be able to distinguish between what the experiment directly demonstrated and what the author is interpreting from it.
That distinction becomes especially important when an AI agent is involved.
Small Editorial Decisions Create Big Improvements
Not every improvement requires sophisticated AI.
Sometimes the biggest gains come from simple editorial changes.
Long headings can be shortened.
A comparison table can receive one sentence of context.
A figure can be explicitly referenced in the surrounding text.
The final sentence of one section can naturally introduce the next.
Background information can be moved closer to the point where it becomes useful.
These changes reduce friction.
The Visuals Can Become the Skeleton of the
One of the strongest ideas in the piece is to begin with the visuals.
Identify the result with the greatest impact.
Move it near the beginning.
Then ask the agent for several ways to visualize it.
Choose the strongest representation.
Repeat the process for the remaining findings.
Once the visual sequence is established, write the narrative around it.
This reverses the traditional workflow.
Instead of writing hundreds of paragraphs and then searching for figures to illustrate them, the figures help determine the structure of the story.
Diagrams Can Carry the Boring Parts
Architecture descriptions and background sections can be difficult to communicate through prose alone.
A diagram can compress several paragraphs into a single visual structure.
The same applies to infrastructure.
Instead of describing a complex experimental pipeline line by line, a diagram can show how data enters the system, how experiments are generated, how models are trained, and how benchmark results feed back into the next iteration.
This makes complicated systems easier to reason about.
The Final Review Still Belongs to Humans
After the agent generates the text and visuals, another review pass is necessary.
Citations need checking.
Numbers need reconciliation.
Transitions need improvement.
Generated-sounding language needs removal.
Claims need to be tested against evidence.
Figures need to be inspected independently.
Finally, the article should be skimmed rather than read word by word.
The question becomes:
If the reader only looked at the title, headings, figures, tables, and a few sentences, would they understand the central argument?
If the answer is no, the article probably needs more editorial work.
Deep Analysis: What This Experiment Really Means
Automated Research Is Becoming a Real Workflow
The most important development here is not simply that an agent found a better data mixture.
It is that the agent was placed inside the experimental loop.
That is a fundamentally different role from generating code on request.
Compute Becomes an Agentic Resource
Once agents can operate research infrastructure, GPU time becomes a resource they can allocate toward hypotheses.
Instead of a human manually deciding every experiment, the agent can determine which experiment to run next based on previous observations.
Data Is Becoming Programmable
Training data is increasingly treated as something that can be manipulated algorithmically.
Mixtures, weights, filtering strategies, deduplication, quality scores, and synthetic-data ratios can all become optimization variables.
Search Can Expose Human Blind Spots
Humans naturally gravitate toward intuitive solutions.
An automated search does not necessarily share those intuitions.
That can be valuable.
A strange-looking mixture may initially seem unreasonable but produce consistently better results.
Benchmark Design Becomes More Important
Automated optimization increases the importance of benchmark quality.
If an agent is repeatedly optimizing against a benchmark, weaknesses in that benchmark can become optimization targets.
A benchmark therefore needs to measure genuine capability rather than superficial patterns.
Broad Improvement Is More Convincing
The fact that the reported mixture improved all ten benchmarks makes the result more compelling than a gain isolated to one evaluation.
Broad improvement provides stronger evidence that the mixture was genuinely useful.
Relative Gains Can Hide Absolute Reality
A 6.8% relative improvement sounds substantial, but readers should also remember the absolute movement from 58.8% to 62.8%.
Both statistics should be reported.
That prevents percentages from becoming misleading marketing language.
More Compute Does Not Automatically Mean Better Research
An 11,000-H100-hour budget is significant.
But throwing compute at a poorly defined objective can produce expensive noise.
The quality of the search strategy matters.
The Agent Needs a Good Objective
An agent optimizes what it is told to optimize.
If the objective is incomplete, the resulting behavior can be incomplete.
Research goals therefore need to be carefully specified.
Evaluation Can Become the Bottleneck
If every candidate requires expensive training and evaluation, even an autonomous agent may face practical limits.
Efficient experimentation remains essential.
Early Stopping Becomes Valuable
An automated researcher should ideally identify weak experiments quickly.
Stopping unpromising runs can preserve compute for more promising hypotheses.
Experiment Logs Become Research Assets
Every automated experiment can produce metadata.
The configuration, training mixture, benchmark scores, runtime, and outcome can form a valuable historical record.
Over time, that history could help agents make better decisions.
Research Agents Could Learn From Their Own Experiments
This is perhaps the most interesting future possibility.
An agent that remembers which data mixtures worked, which failed, and why could become increasingly efficient.
Instead of starting every search from scratch, future experiments could build on accumulated evidence.
Data Mixture Search Could Become Continuous
Rather than selecting a dataset mixture once before training, organizations could periodically reassess mixtures as new datasets and benchmark results appear.
The training pipeline could evolve continuously.
The Researcher Becomes More Like an Architect
As agents handle more implementation details, researchers may spend more time defining objectives, constraints, evaluation frameworks, and interpretation strategies.
The human role moves upward in abstraction.
Editorial Judgment Becomes Scarcer
The
When producing text becomes cheap, deciding what deserves to be said becomes more valuable.
The Same Is True for Visuals
When agents can generate ten charts in seconds, creating a chart is no longer the difficult task.
Choosing the right chart is.
More Options Can Create More Noise
Automation has a downside.
If an agent can generate endless versions of a visualization, researchers can become overwhelmed by alternatives.
Selection remains a human responsibility.
Interactive Articles Could Improve Scientific Literacy
Interactive technical writing can allow readers to explore evidence at their own pace.
That could make complicated research accessible without removing technical depth.
But Interactivity Can Also Become a Distraction
A flashy visualization is not automatically a useful one.
Interactive features should clarify the evidence rather than entertain the reader.
Web-Native Research Could Reduce Cognitive Load
Putting definitions, citations, benchmark explanations, and evidence closer together can reduce the amount of mental context a reader has to maintain.
That is a meaningful advantage.
Static Papers Will Not Disappear
PDFs remain useful.
They are portable, archivable, printable, and easy to preserve.
The future is more likely to involve both static and interactive versions rather than one completely replacing the other.
Agent-Written Research Needs Provenance
Readers should increasingly know which parts of a research workflow were generated, modified, or executed by an agent.
That does not make the work less legitimate.
It makes the process more transparent.
Reproducibility Could Improve
If agents record every experiment, configuration, and result, automated research could potentially create unusually detailed experiment histories.
That could make reproduction easier.
Reproducibility Could Also Become Harder
At the same time, an agent interacting with complex infrastructure may introduce hidden dependencies.
Reproducing the exact environment can therefore remain difficult.
Compute Access Could Become a Competitive Advantage
Organizations with large GPU clusters will be able to run broader automated searches.
This could widen the gap between well-funded laboratories and smaller research teams.
Better Agents Could Reduce That Gap
More efficient search strategies could eventually reduce the amount of compute required to find useful configurations.
The quality of reasoning may become as important as the quantity of hardware.
Data Quality May Matter More Than Data Quantity
A better mixture can sometimes produce more useful training than simply adding more data.
The optimization of composition could therefore become increasingly important as datasets grow.
The Research Loop Is Getting Shorter
The traditional loop is:
Hypothesis → implementation → training → evaluation → analysis → new hypothesis.
Agentic research compresses those stages.
Shorter Loops Can Accelerate Discovery
If experiments can be proposed and evaluated continuously, researchers may explore more ideas within the same calendar period.
That can increase the pace of progress.
Faster Discovery Also Means Faster Mistakes
Automation does not distinguish inherently between productive and unproductive research unless the system is designed to do so.
Bad objectives can be optimized extremely efficiently.
Human Oversight Remains the Safety Layer
Researchers therefore remain essential.
They establish boundaries, inspect unexpected results, challenge assumptions, and decide whether an apparent improvement is scientifically meaningful.
The Future Research Stack May Become Agentic
A plausible future workflow includes an agent for experiment design, another for data analysis, automated infrastructure for training, benchmark systems for evaluation, and agents for producing technical documentation.
Humans would supervise the entire pipeline.
The
The deeper message is that AI is changing both sides of scientific work.
It can help discover results.
It can also help explain them.
The greatest gains may come when those two capabilities are connected.
What Undercode Say:
AI Is Moving From Assistant to Experimental Participant
This experiment represents a meaningful transition from AI-assisted research to AI-participated research.
The agent was not simply asked to write an experiment.
It was given resources and an objective.
The Data Mixture Is an Underappreciated Variable
Model architecture receives enormous attention, but training-data composition can have equally important consequences.
Automating its optimization could become one of the most valuable applications of research agents.
The 6.8% Relative Improvement Is Significant but Needs Context
The reported 6.8% relative improvement sounds impressive, and the underlying four-point absolute increase is meaningful.
However, one experiment should not be interpreted as proof that automated data search universally beats expert-designed mixtures.
Ten-Benchmark Improvement Strengthens the Claim
Improving all ten benchmarks makes the result more persuasive because it reduces the likelihood that the gain was purely benchmark-specific.
Still, independent replication would provide stronger evidence.
Compute Is the Hidden Cost
The experiment consumed 11,000 H100 hours.
That demonstrates capability, but it also highlights the financial and environmental cost of autonomous experimentation.
Efficiency Will Become Critical
The next generation of research agents should not merely find better answers.
They should find them using fewer experiments and less compute.
Research Agents Need Scientific Discipline
An autonomous system can generate hypotheses faster than humans.
That does not mean every hypothesis deserves to be tested.
The best systems will need mechanisms for prioritization.
Benchmarks Will Become Targets
Once agents optimize directly against benchmark scores, benchmark design becomes strategically important.
Weak evaluations could encourage undesirable optimization behavior.
Data Mixture Search Could Become Standard
If repeated experiments confirm these results, automated data-mixture optimization could eventually become a standard stage in model training.
Humans Will Shift Toward Higher-Level Decisions
Researchers may increasingly focus on questions such as what should be optimized, why it matters, and how results should be interpreted.
Editorial Work Is Also Being Automated
The same agentic capabilities can make technical writing dramatically faster.
But faster production increases rather than decreases the need for editorial judgment.
Visual Communication Is Becoming Part of Research
A chart is not merely decoration.
It determines which patterns the reader notices.
Interactive Research Has Real Potential
Web-native technical articles can give readers more context without forcing them to leave the page.
That could become increasingly important for complex AI research.
Fluent AI Writing Is Not Evidence of Correctness
A polished paragraph can still contain an incorrect assumption.
Researchers must verify logic, terminology, numbers, and citations.
Agent-Generated Code and Prose Need Similar Review
The same principle should apply to both.
Trust the output enough to inspect it, but never enough to skip verification.
The Biggest Advantage May Be Iteration Speed
AI’s greatest research contribution may not be one spectacular discovery.
It may be the ability to run and compare hundreds or thousands of reasonable ideas much faster.
More Experiments Change Research Culture
When experimentation becomes cheap, researchers can afford to explore more unconventional hypotheses.
That could create discoveries that would otherwise never be attempted.
But More Experiments Create More Data to Interpret
Automation moves the bottleneck.
Instead of asking how to run an experiment, researchers may increasingly ask which of thousands of results actually matter.
Interpretation Will Become More Valuable
The ability to distinguish signal from noise could become one of the most important research skills.
Research Communication Will Need to Evolve
Technical articles designed for the AI era should exploit interactive media rather than simply reproducing PDF conventions online.
Static and Interactive Formats Should Coexist
There is no need to choose one.
A strong research project can provide an interactive web experience alongside a stable PDF record.
AI Can Help Build the Research Narrative
Once the strongest findings are identified, an agent can help determine the clearest sequence for presenting them.
But the researcher should remain the final editor.
Visual-First Writing Is Powerful
Starting with the strongest figure can reveal the actual story of the research before paragraphs obscure it.
The Best AI Research Workflow Is Hybrid
Humans should provide judgment, objectives, skepticism, and context.
Agents should provide speed, iteration, implementation, and exploration.
The Goal Should Not Be Full Automation
A completely autonomous research pipeline sounds impressive, but scientific progress depends on interpretation and judgment.
A supervised autonomous system may be more useful.
The Future Could Be a Continuous Research Engine
Imagine an agent that continually tests data mixtures, monitors benchmark results, proposes experiments, records failures, and explains meaningful improvements.
That begins to look less like a chatbot and more like a permanent research laboratory.
The Competitive Advantage Will Shift
Organizations may compete not only on models and GPUs but on the quality of their automated research loops.
The best research infrastructure could become a strategic advantage.
Autoresearch Is a Concept Worth Watching
The experiment is an early indication of what happens when AI agents receive both objectives and resources.
Its long-term importance will depend on replication, efficiency, generalization, and scientific reliability.
✅ Confirmed: The experiment described a budget of approximately 11,000 H100 GPU hours for searching for a better data mixture for a 2B vision-language model.
✅ Confirmed: The reported best mixture increased the aggregate score from 58.8% to 62.8%, equivalent to roughly a 6.8% relative improvement.
✅ Confirmed: The reported optimized mixture improved performance across all ten benchmarks evaluated in the experiment.
❌ Not established: These results alone do not prove that autonomous agents will consistently outperform expert-designed data mixtures across models, datasets, or research environments.
❌ Not established: The experiment does not demonstrate that human researchers can be removed from the research process; objective design, evaluation, interpretation, and verification remain critical.
Prediction
(+1) Autonomous Data Optimization Will Grow
As research agents become better at planning and executing experiments, automated data-mixture search is likely to become a more common component of model development.
(+1) Research Agents Will Run Larger Experimental Loops
Future systems will increasingly propose experiments, execute them, analyze the results, and select the next experiment with limited human intervention.
(+1) Data Composition Will Become a First-Class Optimization Problem
Researchers are likely to devote more attention to how different datasets interact rather than treating dataset construction as a mostly manual preprocessing stage.
(+1) Interactive Research Articles Will Become More Common
As agents make visualization and web development cheaper, more technical research will move toward interactive presentations that combine prose, figures, demonstrations, citations, and explorable data.
(-1) Compute Costs Could Limit Adoption
Large-scale autonomous experimentation may remain concentrated among organizations with substantial GPU resources unless research agents become dramatically more compute-efficient.
(-1) Benchmark Gaming Could Become a Bigger Problem
If agents are rewarded primarily for benchmark scores, researchers may face increasing pressure to design evaluations that resist narrow optimization and better represent real-world capability.
(+1) Human Researchers Will Become More Like Research Directors
Rather than disappearing, researchers are likely to spend more time defining questions, setting constraints, evaluating evidence, and deciding which discoveries actually matter.
(+1) The Biggest Long-Term Advantage May Be Research Speed
If agents can compress weeks of experimental iteration into days or hours while maintaining scientific rigor, the resulting increase in research velocity could become one of the most important advantages of agentic AI.
(+1) The Research Lab of the Future Will Be Partly Autonomous
The direction suggested by this experiment is clear: the next generation of AI laboratories may combine human judgment with continuously operating agents that search, test, analyze, visualize, and document new ideas.
The tools are becoming faster.
The experiments are becoming more automated.
And as the cost of producing research continues to fall, the most valuable skill may increasingly be knowing which questions are worth asking in the first place.
▶️ Related Video (76% Match):
🕵️📝Let’s dive deep and fact‑check.
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
References:
Reported By: huggingface.co
Extra Source Hub (Possible Sources for article):
https://www.discord.com
Wikipedia
OpenAi & Undercode AI
Image Source:
Unsplash
Undercode AI DI v2
🔐JOIN OUR CYBER WORLD [ CVE News • HackMonitor • UndercodeNews ]
📢 Follow UndercodeNews & Stay Tuned:
𝕏 formerly Twitter 🐦 | @ Threads | 🔗 Linkedin | 🦋BlueSky | 🐘Mastodon | 📺Youtube
