<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ignacy Ruszkowski on Medium]]></title>
        <description><![CDATA[Stories by Ignacy Ruszkowski on Medium]]></description>
        <link>https://medium.com/@ignacyrusz?source=rss-95a129948399------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*5MaeGqKAU9AmJ5oDV2IR1w.png</url>
            <title>Stories by Ignacy Ruszkowski on Medium</title>
            <link>https://medium.com/@ignacyrusz?source=rss-95a129948399------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 22:25:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ignacyrusz/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[This Paper Changed How I Think About Risks of Open-Weight Models.]]></title>
            <link>https://medium.com/@ignacyrusz/this-paper-changed-how-i-think-about-risks-of-open-weight-models-fcd459ef8853?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/fcd459ef8853</guid>
            <category><![CDATA[research]]></category>
            <category><![CDATA[summary]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[risk-management]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Sat, 12 Sep 2026 18:35:07 GMT</pubDate>
            <atom:updated>2026-09-12T18:35:07.988Z</atom:updated>
            <content:encoded><![CDATA[<h3>Open-Weight Models Are More Dangerous They We Thought.</h3><p>This is a summary of Casper et al. (2026)<a href="https://openreview.net/challenge?redirect=%2Fforum%3Fid%3D8QyGLnFkzc"> Open Technical Problems in Open-Weight AI Model Risk Management</a> paper, along with some of my commentary, pointing out which points made me change how I think about open-weight models and their safety.</p><h3>TL;DR</h3><p>Open-weight models are getting better day by day, with the state of the art models trailing roughly <a href="https://epoch.ai/data-insights/open-closed-eci-gap">four months</a> behind the proprietary models. These models pose new challenges, as they can:</p><ol><li>be modified arbitrarily,</li><li>be used without oversight,</li><li>spread irreversibly.</li></ol><p>Authors point out 16 open technical challenges for increasing safety of open-weight models. My main takeaway from this paper: we need a higher standard of risk management from companies releasing open-weight models. Their current approach poses societal risks which could be decreased if more effort was put towards it.</p><h3>Introduction</h3><p>With open-weight AI models usage growing quickly, and their capabilities lagging four months behind frontier proprietary models, they likely possess dangerous capabilities already. The deployment of open-weight models differs from the closed-weight ones, with the former having a bigger chance to be misused. Methods which are used to secure proprietary models can’t be utilized for open-weight models, whose attack surface is larger and toolkit for defense techniques is less established.</p><p>Work on technical safeguards for open models is critical to reap their benefits while offsetting the associated risks. The key areas in which open problems are stated are:</p><ol><li>Training data curation, i.e. providing data which doesn’t contain harmful content,</li><li>Tamper-resistant training and “unlearning” algorithms,</li><li>Model tampering evaluations,</li><li>Staged deployment strategies,</li><li>Model provenance and forensics.</li></ol><h3>The challenges of open-weight models</h3><p>The most interesting part of the paper to me was the realization that the challenges associated with open-weight models are very different from those of their closed-weight counterparts. The work also increased my awareness of unique challenges posed by open-weight models, presenting how they are way harder to maintain under control than proprietary models. Managing the risks of a model is more problematic if the model can be replicated, modified and deployed by almost anyone.</p><p>Open-weight models are double-edged swords. Their openness is key to enabling more widespread research about them and decentralization of power, but creates unique problems. For example, users can get rid of safeguards that hamper misuse, a method also called tampering.</p><p>Open-weight models spread rapidly and once their weights are published, they can’t be retrieved. Another downside of these models is that their use can’t be centrally monitored or moderated, as they are spread over independent machines with no authority governing their use. This prohibits ‘Know Your Customer’ policies, enabling everyone to harness the model’s power without any supervision. This is concerning.</p><p>The supply chains of open-weight models are also more complex than closed-weight ones, they do not have a centralized core and the talent, resources, data, compute and infrastructure is more distributed than that of the closed-weight models. With this in mind, it’s also harder to track real-world spread, usage and impact of open-weight models.</p><p>As an addition, it is worth mentioning that <a href="https://x.com/DarioAmodei/status/2098773920774074715?s=20">Dario Amodei</a> of Anthropic and <a href="https://x.com/sama/status/2098811563415150910?s=20">Sam Altman</a> of OpenAI very recently expressed interest in increasing safety of their models by enabling embedded evaluators to look into the company’s internals as if they were their employees. This could possible enhance the security of proprietary models, it is interesting to see whether a similar idea will come spread to the Chinese labs as well.</p><h3>Toolkit</h3><p>The emphasis of the paper is put on open problems for technical safeguards with distinct implications for open-weight models. A taxonomy is proposed to manage the risks of open-weight models. Tools that do not have direct implications on open-weight models are not considered in this paper, neither are nontechnical approaches, governance strategies nor development of open-weight models with benefits for society.</p><h3>Open Technical Problems</h3><p>The first category is <strong>Training Data Curation</strong>, which is championed as a key intervention for improving model safety. If a model is not exposed to harmful data during training the risks of it having propensity for such behaviors are minimized. Not all current approaches focus on preventing harmful content from entering training data.</p><blockquote>How does data curation’s effectiveness differ across harm categories?</blockquote><p>Recent work in this direction is promising. Pretraining filtering’s ability has been shown to reduce model’s knowledge of biorisk-related topics.</p><blockquote>How can scaling data curation be scaled across languages, modalities and data/model sizes?</blockquote><p>Given the size of internet-scale datasets the feedback loops might be on the longer side. Previous approaches traded-off efficiency for precision, with many good documents being cut off. It’s not clear whether bigger and more sample-efficient models need relatively more filtering.</p><blockquote>What is the relationship between training data contents and emergent model capabilities?</blockquote><p>Can dangerous capabilities emerge from neutral data in practice? Can information from one domain have a bad effect on data from another domain?</p><p>The second category is <strong>training which is resistant to modifying model’s downstream capabilities and unlearning algorithms</strong>. Resisting such attacks could improve safety under adversarial manipulations. Post-training safety guardrails are key in this strategy. State of the art (SOTA) models have been vulnerable to such modifications in the past.</p><blockquote>How do we develop more tamper-resistant unlearning algorithms?</blockquote><p>Leveraging mechanistic interpretability in the context of topics about which models are ignorant could help with enhancing training objectives. There might, however, be limitations for post-training methods’ abilities to remove unwanted knowledge from models.</p><blockquote>How can we robustly edit model beliefs with minimal side effects?</blockquote><p>Instead of making models ignorant about certain topics, introducing certain incorrect beliefs about hazardous topics could counteract harmful output generation. For example models would learn an <strong>in</strong>correct way to obtain abusive material.</p><blockquote>Can we develop models that effectively resist retrieving harmful information?</blockquote><p>With access to tools, models can learn harmful information without having it in its memory, by browsing the web or other files. Preventing this from happening remains an open question.</p><p><strong>Evaluating models which are trained to have safeguards removed downstream.</strong> It’s important to assess practical risks from open-weight models that have been modified. Currently, such evaluations are not a standard practice.</p><blockquote>How can we develop rigorous benchmarking and evaluation frameworks?</blockquote><p>The risks are widely understood, but their occurrence is not reported often enough.</p><blockquote>What modifications should be used for worst-case risk estimation under model tampering?</blockquote><p>Threats other than fine-tuning ones, like low-rank modifications, latent-space attacks or distillation can impair safety of models, and should be studied more.</p><blockquote>How can we systematically identify effective attacks and defenses?</blockquote><p>Tampering attacks are computationally intensive, developing practices to do such evaluations at a scale, while remaining rigorous would require lowering the compute cost.</p><blockquote>How can we scalably evaluate thousands of models?</blockquote><p>The open-weight ecosystem includes a huge number of models. Assessing their safety at scale would improve practical risk management to a big extent.</p><p><strong>Staged deployment strategies</strong> are the next theme that is talked about. Slowly rolling out models before fully open-sourcing them would help developers monitoring risks and adjust safeguards as well as deployment strategies.</p><blockquote><em>What exfiltration risks do split deployment strategies pose, and how can they be mitigated?</em></blockquote><p>Private model layers would need to be kept secure for successful deployment, but attackers can try exfiltrating the private layers through reconstruction.</p><blockquote>How can we design split learning and inference APIs that are less costly and more competitive?</blockquote><p>Communication across the split is hard and introduces latency. The overhead is a limitation to other deployment forms, making split strategies less competitive.</p><blockquote>Can hardware locking or homomorphic encryption offer practical options for staged deployment?</blockquote><p>Hardware locking introduces security guarantees within physical processors themselves. In machine learning this typically means binding model weights to a secure, isolated area of a CPU or a GPU. Homomorphic encryption enables computations to be done on encrypted data without ever deencrypting them. Both of these methods could improve the security of open-weight models, but it remains unclear in what way they would do so. It’s not sure if they would find broad applications.</p><p>The last category that the paper talks about is <strong>model provenance and forensics.</strong> Both provenance and forensics help people examine the spread and use of open-weight models, and track the lineage of what the base model was, of a particular fine-tuned version.</p><blockquote>How can we watermark models in ways that are more durable against common modifications without side effects?</blockquote><p>Evaluations suggest that current solutions in this space can be erased through quantization, fine-tuning, model merging and pruning. Lack of standardized benchmarks that compare watermark durability is also a problem.</p><blockquote>What algorithms can enable scalable and versatile model heritage inference?</blockquote><p>It would be desirable to track the heritage of models across the whole ecosystem, but with current infrastructure and methods it is not possible.</p><blockquote>How practical and scalable are proof of training methods?</blockquote><p>Proof of training methods refer to ways to formally verify whether a specific neural network was genuinely produced by a particular computational training run, whether the model was or wasn’t trained on a given dataset or if the training adhered to safety, governance and algorithmic constraints.</p><h3>What techniques are open-weight developers reporting on?</h3><p>The results of this section stem from model cards of popular open-weight models. In short, almost none of the solutions mentioned above are being talked about in most popular system cards. Data curation stood out as the most common technique but reports on the rest of the techniques are almost non-existent. There is a lot of potential in increasingly reporting measures talked about in this paper.</p><p>This table is one of the most interesting findings in the paper. It shows how almost none of the open problems are mentioned in the models’ system cards.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5j7wVC3F-duVF8T9.png" /></figure><p>The table above portrays a big gap in the reporting of open-weight models. Open-weight creators do not report the associated risks clearly, most likely omitting ever measuring them.</p><h3>Discussion</h3><p>Open-weight model capabilities remain four months behind frontier proprietary models. This is not a lot. They pose specific risks due to their fast-spreading nature and ease of which their safeguards can be bypassed. The paper focused only on technical tools, but a more general monitoring and risk mitigating approach is crucial for capturing the benefits of this technology. The extent to which safeguards for open-weight models will ultimately be is unclear, but even with safeguards some models might pose serious risks.</p><p>Gathering more information about open-weight models is valuable, and in doing so the community should be aware of ‘openness washing’ as well as ‘safety washing’, analogies to the well known greenwashing phenomenon.</p><p>Current incentives are not well aligned for private actors to research and develop safeguards for frontier open-weight models. Even if they were, safeguards are effective only to some extent, therefore working on this issue is potentially a high-risk and limited reward undertaking.</p><p>Authors end by emphasizing the value of open scientific collaborations, open evaluations, open reporting about risk-management methodology and open standardized documentation.</p><p>Through engaging with this paper I changed my perspective on how risky open-weight models are. Current releases lack the sophistication and rigidness of necessary safeguards reporting. In the light of the recent commitments to embedded evaluators from Anthropic and <a href="https://x.com/sama/status/2098811563415150910?s=20">OpenAI</a>, I would like to see more care put into the state of open-weight models as well.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fcd459ef8853" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Being Second Best In the AGI Race Just Got Way Cheaper]]></title>
            <link>https://medium.com/@ignacyrusz/being-second-best-in-the-agi-race-just-got-way-cheaper-b2a57b016e45?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/b2a57b016e45</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[china]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Sun, 26 Jul 2026 13:01:02 GMT</pubDate>
            <atom:updated>2026-07-26T13:01:02.009Z</atom:updated>
            <content:encoded><![CDATA[<p>How Distillation Makes This Possible And How The Race Might Change.</p><h3>Kimi K3 release was special.</h3><p>Last week, on July 16th <a href="https://kimi-k2.org/kimi-k3">Kimi K3</a> was released by the Chinese company Moonshot AI. It took the AI world by a storm, and rightfully so. The model pushed the frontier of open-weight models (the weights have not been published as of July 26th 2026, they are scheduled to be released on July 27th), almost matching other SOTA models, while being available at a significantly lower cost than Fable 5. All of this was achieved among accusations of using a method that gathered a lot of attention recently — distillation.</p><p>While not officially disclosed, there are several cues for <a href="https://stephen.bochinski.dev/blog/2026/07/18/the-kimi-k3-moment/">Kimi K3 being trained on outputs of Claude</a> including K3 identifying itself as <a href="https://github.com/rgreenblatt/which_claude_is_k3/blob/main/writeups/write_up.md">Claude 15% of time</a>. People have gone on to compare the <a href="https://typebulb.com/u/lab/you-re-relatively-right/full">similarity of answers to the same prompts across models</a> and have seen similar results. White House OSTP Director <a href="https://www.reuters.com/world/us/us-accuses-chinas-moonshot-stealing-anthropics-fable-latest-ai-model-2026-06-30/">accused Moonshot AI of distilling Kimi K3 from Fable 5</a> , but did not provide any proof to the public. Back in February 2026 <a href="https://www.reuters.com/world/china/chinese-companies-used-claude-improve-own-models-anthropic-says-2026-02-23/">Anthropic claimed that Chinese AI companies distilled Claude</a> to train their models. While not confirmed, it is plausible that K3 was trained on Claude’s data. Kimi addressed the accusations pointing at the timeline, there were 37 days between the release of Fable 5, and Kimi K3, which the company claims to be too short to train a state of the art model.</p><h3>This method is called distillation, because data is distilled from models who have learned it from other sources.</h3><p>The method involves pre-training models on data which was sourced from other models instead of the internet, which would likely yield worse results. In the past, we expected that <a href="https://www.nature.com/articles/s41586-024-07566-y">using synthetic data recursively lead to model collapse</a>, an Ouroboros snake eating its own tail approach doesn’t give good results when training language models. But those results focused on the data’s recursive nature, and in the distillation as we think of it now we are dealing with curated synthetic data.</p><figure><img alt="Ouroboros, a snake eating its own tail. An ancient Egyptian symbol of constant revival." src="https://cdn-images-1.medium.com/max/1024/0*OmD8oxlPQndrc64C.png" /><figcaption>Ouroboros, a snake eating its own tail. An ancient Egyptian symbol of constant revival.</figcaption></figure><p>Ouroboros, a snake eating its own tail. An ancient Egyptian symbol of constant revival.</p><p>But a shift has taken place, and if you consider what is happening in reality it does make sense. If you train your models on data scraped from the internet, a <a href="https://mynewitguys.com/what-percentage-of-online-content-is-ai-generated-in-2025/">huge part of it is already synthetic</a>, i.e. created by bots. With state of the art language models reaching performance of <a href="https://artificialanalysis.ai/evaluations/humanitys-last-exam">PhD students or higher</a> across multiple disciplines, it is easy to see why they will be of higher quality than an average post on X or Reddit.</p><figure><img alt="Performance of different models across time od GPQA Diamond accuracy benchmark, an AI performance test on PhD level science questions." src="https://cdn-images-1.medium.com/max/1024/0*WdjjDlO_9eQB8Cmb.png" /><figcaption>Performance of different models across time od GPQA Diamond accuracy benchmark, an AI performance test on PhD level science questions.</figcaption></figure><p>Performance of different models across time od GPQA Diamond accuracy benchmark, an AI performance test on PhD level science questions.</p><h3>This is unethical! We can’t let this happen!</h3><p>Many have raised concerns about the ethics of this tactic, which violates the company’s terms of service. It is nothing short of piggybacking on frontier labs’ research to train your models and releasing them for a fraction of the price feels especially unfair. While in an ideal world, this wouldn’t happen, in reality this approach doesn’t seem any more <em>unethical</em> than scraping the whole internet, without having any permission for it whatsoever. Lawsuits like <a href="https://harvardlawreview.org/blog/2024/04/nyt-v-openai-the-timess-about-face/">The New York Times Co. v. Microsoft Corp. and OpenAI</a> are possibly only by huge efforts and use of capital by the biggest companies, independent creators have no way of hoping to have any chances at such a lawsuit, which would protect the tech giants from profiting off their work.</p><p>More than anything else, this seems to be an incentives problem. Currently, companies providing best quality outputs for cheap have a great shot at leading the race, and this is made possible by significantly cutting the training costs by stealing other’s results. Majority of the AI bubble doesn’t care about the ethics and factors other than price and quality of output these days.</p><h3>People were amazed by Kimi K3.</h3><p>While claiming that K3 is the best open-weight model would be too optimistic, given the amount of different benchmarks out there, it is not far from the truth, with the performance across multiple metrics, like <a href="https://artificialanalysis.ai/agents/coding-agents?coding-agents-performance-chart=deep-swe">coding</a> and placing high in the <a href="https://artificialanalysis.ai/leaderboards/models">Artificial Analysis’ LLM Leaderboard</a>, behind only GPT and Claude models as of July 26th 2026. While this release gathered a lot of attention from the general public, I don’t think the distillation point was covered enough. I believe it has a possibility to reshape the race towards AGI tremendously.</p><h3>A day before K3 was released, Thinking Machines revealed Inkling.</h3><p>This was celebrated as the best American open-weights model to be released as of recently. It was cheap, and claimed to be a generalist model being able to work with audio. Thinking Machines claims to have trained the model from scratch, and there doesn’t seem to be any evidence suggesting that it could have been distilled. The hype around it quickly faded as K3 came into the picture, as its performance just wasn’t that special.</p><h3>State of open-weight models and how the dynamics of frontier labs have to change.</h3><p>What didn’t change is that to be the company with the best models publicly available is very costly. The novel thing is that being the company with the second-best model out there has gotten significantly cheaper1 with distillation, if you manage to train your models fast enough you can be only months to weeks behind the leading AI labs, at a fraction of the cost. This is especially true if the U.S. jurisdiction doesn’t concern you too much, but even there some of the biggest tech companies, including Nvidia, Microsoft, IBM and OpenAI, have gathered to <a href="https://news.ycombinator.com/item?id=49035303">prevent banning open-weights models</a>. In other words, the incentives to be the first might diminish, because it is so much cheaper to be trailing. The costs do not sound like a real concern for the biggest companies in the world, but if you consider the pressure they face to produce returns for their capital investors, it is a real issue. Companies like Meta, Oracle, Amazon, Microsoft and Alphabet (owner of Google) have been accused of <a href="https://archive.ph/wlbqs">hiding their debt from the balance sheet</a> to pour more money into investment.</p><h3>What’s coming next?</h3><p>We would expect to see stronger guardrails embedded into models that protect against distillation. Frontier companies have already put a lot of effort into preventing the usage of their models to develop other models using them. So far, these approaches proved to be a double edged sword, as the models with stronger guardrails were deemed less usable. For example Fable 5 had guardrails that often “routed” its users to a less capable and more protected Opus 4.8 when it believed the users are trying to do something illegal, like perform cyber attacks or develop bioweapons. While the intent was sound, the application was dubbed underdeveloped, as it prevented developers from using the model’s capabilities to perform any serious work in cybersecurity and other use cases the model flagged as dangerous. In the days of distilled models which do not exert so much effort in terms of safety, people can simply switch to somewhat less capable but more usable models like Kimi K3. This is once again a problem of incentives forming in a bad way, people are more likely to use the <em>dangerous</em> models because they enable them to do the <em>safe</em> work they are trying to do. This is bad news for AI Safety.</p><h3>The AGI race might finally slow down.</h3><p>It might look like companies will take more time to develop appropriate guardrails before releasing their frontier models, or else they will feed their competitor’s growth. Like in competitive road cycling, the rider in the front exerts the most effort and is unlikely to finish first, it is more likely for the people trailing just behind him to overtake at a crucial moment and win instead. While judging whether the race actually slows down will take at least a few months, Anthropic releasing <a href="https://www.anthropic.com/news/claude-opus-5">Opus 5 just two days ago</a> and claiming it to be less restrictive and more affordable than Fable 5 signals that the labs notice their shortcomings.</p><p>Another plausible scenario is that the gap between internal models (with fewer guardrails) and external ones, available to the public (with guardrails on) might grow bigger, and we won’t learn about it until a scandals like the <a href="https://www.nytimes.com/2026/07/21/technology/openai-attack-hugging-face.html">Hugging Face OpenAI hacking</a> come around. This is, once again, not a good thing for the overall safety, as external auditing will be severely limited if not impossible for models which are gatekept within the AI labs.</p><p><strong>Footnotes</strong></p><ol><li>Inferred based on lower usage costs.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b2a57b016e45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Europe Must Build Sovereign AI Or Risk Irrelevance.]]></title>
            <link>https://medium.com/@ignacyrusz/why-europe-must-build-sovereign-ai-or-risk-irrelevance-8fa412ce320f?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/8fa412ce320f</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[politics]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 09:31:53 GMT</pubDate>
            <atom:updated>2026-07-15T09:33:15.088Z</atom:updated>
            <content:encoded><![CDATA[<h3>Remember AI 2027?</h3><p>The researchers behind that project are back with another great piece, expanding their visions about possible futures with AI developments, available at <a href="https://markdowntorichtext.com/ai-2040.com">ai-2040.com</a>. Instead of forecasting the future, they propose several alternative paths, indicating which ones are desired. This time around, the team proposes several parallel timelines with each of them changing at key decision points. I went through it. It made me think about geopolitics. I became increasingly aware of the role of Europe in the advertised scenario, dubbed Plan A. The thing is, Europe’s role is close to non-existent. Let me explain why is this the case and what has to change about Europe’s approach if people across the whole world are to benefit.</p><h3>But first a brief recap of ai-2040.</h3><p>The research outlines different plans, A, B, C, D and S, and their outcomes. The timelines diverge in 2029, when an important decision has to be made. The scenarios stretch all the way to, as you might have guessed, 2040. Plan A, i.e. the plan that the website champions as the best scenario is the <em>Verified Slowdown</em>. Plan B is slowing down and forcing China to do so as well (as U.S.), C is racing and using the advantage to develop safety measures, etc. Let’s focus on Plan A, as that is the path most vigorously discussed, and it is proposed as the <em>best</em> solution. The timelines are presented mainly from the perspective of USA, which is fair given that the majority of AI development is currently concentrated in the USA, and the USA is, well, the USA.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U4n9GedvHi5WU6AW.png" /><figcaption>Paths proposed by ai-2040. Plan A is the recommended option.</figcaption></figure><p>ai-2040 talks about the existence of two workforces, working in parallel, one of them being humans and the other one being AI agents. It acknowledges that current AI Agents capabilities are still limited, but it predicts they will get better and increasingly become a more and more important part of the global economy. In particular, they will help the economies that have access to AI agents powerful enough to do any meaningful job. Under Plan A, U.S. and China become the two most important players in the AI race and they get to decide what the rest of the world has to agree to. They get to decide the future. In the wishful scenario they agree to a form of “Mutually Assured Compute Destruction” (MACD), in which they don’t have to rely on mutual trust, but on performing checks instead. This agreement is largely based on the <a href="https://www.britannica.com/topic/mutual-assured-destruction">MAD</a> strategy composed for deterring the use of weapons by retaining as much weapons to destroy the enemy, while the enemy holds onto the amount they need to destroy you. This approach acknowledged that complete destruction of nuclear weapons was not realistic, but there were paths to safety. AI safety research draws a lot of parallels with MAD.</p><p>What was striking for me in this scenario is the lack of virtually any role for Europe in it. It got me thinking…</p><h3>Let’s talk about Claude Mythos for a second.</h3><p>This is an important piece of the puzzle. Anthropic announced Claude Mythos on April 7, 2026 claiming it to be a massive upgrade in existing model capabilities. At the same time, the company made it clear that this model will not be released to the general public due to the security risks it carries. They literally claimed their model carried too many cybersecurity and biology misuse risks for them to release it, and they decided to the access to it to chosen companies anyways. After the hype died down a little, <a href="https://www.schneier.com/news/archives/2026/04/claude-mythos-is-mostly-marketing-hype.html">experts</a> from the industry expressed doubts whether the model was as capable as the company claimed, and some people said that the <em>gated for safety</em> release was a <a href="https://daniel.haxx.se/blog/2026/05/11/mythos-finds-a-curl-vulnerability/">marketing stunt</a>. Whatever it was, it portrays an important dynamic. When a newly built model is very capable, it might be released only to a number of hand-picked companies and they will be the ones to reap the benefits. Importantly, Anthropic disabled their most powerful Fable 5 and Mythos models after a U.S. government export-control directive ordered it to stop access for foreign nationals. This shows even AI companies are not above the law and governments can certainly make them bend.</p><h3>Who gets to decide the future?</h3><p>It increasingly seems like the access to the most capable AI models will be an important factor in shaping the economies of tomorrow, and the people responsible for granting or restricting this access will exercise a great power in deciding who gets to grow and who will have to shrink instead. The perspective presented by ai-2040 shows it clearly. The markets might shift to be an outcome of automated work performed by AI agents and robots, in which scenario having access to them matters a lot. Naturally, the power connected that comes with AI dominance concentrates in the U.S., as of now.</p><p><a href="https://www.youtube.com/@hoog-youtube">Hoog</a> is a Dutch creator which recently presented an important European perspective that inspired me to think about the issue of AI access across the globe. In their recent <a href="https://www.youtube.com/watch?v=C84Lnyv0gTs">video</a>, Hoog presents a scenario that is bound to happen unless Europe does something remarkable. They present a scenario similar to the one of ai-2040, but give it a different framing. It is their perspective that prompted me to think about these questions seriously.</p><h3>AI champions across the globe.</h3><p>The landscape is changing rapidly but there are a few companies which managed to stay on top for the past few years in the capabilities race. Most notable companies working on AI in the U.S. are OpenAI with ChatGPT, Anthropic with Claude family of models, xAI with Grok, Google with Gemini, Meta with Llama family and plenty more companies whose models are less capable but could become influential in the future. They certainly have their champions, and that includes not only the companies, but also talent, data centers and the whole infrastructure around it. China is slowly closing in on the leaders of the race with their champions including DeepSeek, Zhipu ai (known by it’s website Z.ai, the company behind GLM 5.2), Alibaba with Qwen, MiniMax with their models, Moonshoot AI behind Kimi assistant and many more examples. It suffices to say that these two countries have their say in the AI space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nXS0Q9-CVX7e9QXp.png" /><figcaption>The U.S. and China have no incentives to help Europe. The continent has to help itself.</figcaption></figure><h3>Europe doesn’t compare.</h3><p>What’s left in Europe is surprisingly little. Firstly, Mistral is based out of France and remains the most capable European foundational model, that is besides DeepMind’s work whose interests after the acquisition by Google I consider American. OpenAI has thus far raised about <a href="https://tracxn.com/d/companies/openai/__kElhSG7uVGeFk1i71Co9-nwFtmtyMVT7f-YHMn4TFBg">$180B</a> over it’s funding rounds, and Mistral has pocketed <a href="https://tracxn.com/d/companies/mistral-ai/__SLZq7rzxLYqqA97jtPwO09jLDeb76RVJVb306OhciWU/funding-and-investors">$3.05B</a>, roughly 60 times less, while being the biggest model out of the continent. Comparing funding is not the best metric for model’s capabilities, but if you look at the sheer number of AI companies coming out of the U.S. and China as well as the capability benchmarks like <a href="https://arena.ai/leaderboard/code/webdev">Code Arena</a> the situation is quite obvious (Mistral’s highest ranked model is 80th at the time of the writing of this post, with U.S. and China ahead). Europe has a lot of catching up to do. In the next paragraphs I will explain why this is bad thing not only for the people from Europe but for all of us, what levers can Europeans pull to change their course and what could happen as a result.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*76qXwrJ7aroj0ME7.png" /><figcaption>Comparison of Mistral AI (leading European foundational model company) and OpenAI funding. Funding is not the a direct measure for model capability.</figcaption></figure><h3>Why would anyone outside of Europe, care about Europe?</h3><p>We live in a world which is very interconnected, and goods and services can flow through continents’ boundaries. Importantly, international cooperation happens because of legislations and governance is in place to enable them. Europe has been a leader that adopted multiple innovative rules which propagated through the legislation systems of the world. It put liberal democracy and parliamentary institutions as a benchmark, while actively supporting election observation and democracy-building abroad. The continent’s European Union showed that former rivals can be reunited by cooperating between their markets and institutions, creating peace as a result. European privacy and platform accountability rules have had impact on numerous companies and regulators across the world, because internationals tend to adopt the strictest standard across markets, which happened to be Europe. There was also a lot of climate and environmental initiatives that drove climate diplomacy and standards across the world, helping enforce rules on emissions, sustainability and green industry.</p><h3>Europe can bounce back.</h3><p>The most important asset the old continent currently has is its largest company by market cap, the Dutch extreme ultraviolet (EUV) lithography machines producer, ASML. It is the only company capable of producing machines necessary for modern semiconductor industry, and by choosing who the company sells their machines to, has a huge influence on the global markets. Importantly, the key talent is also within Europe with top-tier technical universities like Imperial College London, ETH Zurich, Technical University of Munich among the leaders, it has the academic institutions to attract the best researchers in the world. With necessary revisions of how research is done at scale in those universities (particularly by spreading the model of universities like ETH or EPFL to allow professors to supervise more student efficiently) and appropriate grants coming from the governments and EU, it could catch up by a lot. Innovative environments which were created among the leading universities, with the most notable startup hubs in London, Paris and Munich could rival their oversees rivals, if perhaps more money was made available to them. The money needed for financing innovation is in Europe, but being held in pensions funds, it is discouraged from being invested in risky ventures. Proper regulations would need to be introduced to treat these problems.</p><h3>Criticizing is always easier than giving helpful feedback.</h3><p>People that have the power to change things often get stuck in negative loops that make them ramble about how stuff doesn’t work and how they wish somebody finally fixed it, never realizing that they are in power to do exactly that. Capabilities of AI models increase, more and more could be gatekept. With these changes, there might finally be a rally around the flag effect in Europe which will make its residents stand up, take a stance and get to work instead of arguing. Let’s hope we don’t have to wait for dramatic loses before a breakthrough comes around, making people gather and cooperate.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8fa412ce320f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Axioms and Assumptions Are Made Up. ]]></title>
            <link>https://medium.com/@ignacyrusz/axioms-and-assumptions-are-made-up-4ae0a5f6d015?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ae0a5f6d015</guid>
            <category><![CDATA[assumptions]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[axioms]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Tue, 07 Jul 2026 11:06:12 GMT</pubDate>
            <atom:updated>2026-07-07T11:06:12.703Z</atom:updated>
            <content:encoded><![CDATA[<h3>Axioms and Assumptions Are Made Up. How We Invented Paradigms To Change The World.</h3><h3>Paradigms are sets of rules accepted by groups of people working in a particular field.</h3><p>Scientists agree to them while solving problems deemed important and worthy of a solution. Paradigms, comprised of axioms, i.e. things we hold as true, arise among scientific communities focused on a particular area or a problem. If you think about doing science or research as puzzle solving, the are the rules along which you agree to play, like you can only join puzzles together if they match with each other and can be pieced together effortlessly. They also talk about things one cannot do within a particular paradigm, like force the puzzles into a particular position.</p><p>A paradigm does not have to be accepted by all scientists working in a particular domain. Instead, it is common to see numerous subgroups of specialists within a field agreeing upon different axioms and working under numerous paradigms.</p><p>When a group of scientists including Claude Shannon (the founder of Information Science) and Marvin Minsky (one of the founding fathers of AI) gathered together in the Dartmouth Summer Research Project in 1956 didn’t expect they <em>will not have solved AI</em> by the time Autumn comes around. Obviously, they did not anticipate the emergence of two separate schools of doing AI neither, with the divide between symbolists and connectionists. The disputed between the two could be traced down to a disagreement of what intelligence really was. The former group believed that by setting rules and encoding symbols one can create intelligence, while the latter believed learning from experience, by running training on data and letting computers figure out a way, is better. It is the connectionists that emerged victorious (thus far) and it is their approach that lures behind modern machine learning approaches including neural networks and deep learning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*P3njlbsmqWt6gvzp" /><figcaption>Dartmouth 1956 Summer Research Project Members</figcaption></figure><h3>Paradigms enable groups to look at the same scenario and see the same things happening.</h3><p>If you show five distinct connectionists how a trained neural network tells apart an image of a handwritten letter P from the letter R, they will say that what you see is an emerging intelligence. On the other hand if you show the same thing to the symbolists, they will say that it is just a bunch of random numbers predicting something, not even remotely close to an artificial intelligence.</p><p>Comparing two different paradigms and deciding which one is better with certainty is not possible. While it might seem that rules are more accurate versions of others, it’s hard, if not impossible, <em>to prove</em> that one of them is superior. This is especially the case if they are distinct, like Newtonian and quantum mechanics. It is so because they tend to differ on fundamental questions and because they invariably take some axioms as true. Thus, they will have a hard time proving their case to people from another school which agrees upon other axioms instead. From this distinction emerges another important fact about paradigms.</p><h3>Because people operating within a paradigm accept certain things as true, they also agree about what questions are important to solve and which ones are not.</h3><p>If a hypothetical scientist accepts the ability to predict stuff as a sign of intelligence, he will not try to question this narrative, but instead move on to answer the question of how good certain predictions are and distinguish different types of intelligence.</p><p>Paradigms tend to have questions that they aim to solve while accepting answers to other questions as given.</p><p>This approach also proves quite effective, i.e. it allows researchers to focus on what they have agreed <em>that matters</em> and neglect other things. If a paradigm is stable, science progresses by answering questions with expected possible solutions. As Thomas S. Kuhn the author of <em>Structure of Scientifc Revolutions</em> says in his book:</p><blockquote><em>As in manufacture so in science — retooling is an extravagance to be reserved</em></blockquote><p>Most science progresses when paradigms are not questioned, because they are accurate enough for the scientists to pursue their investigations. Research under stable paradigms resembles the act of puzzle solving. People expect, more or less, what they should find, just like puzzle-solvers have an idea of what the finished puzzle will look like. Their paradigms supply them with theories that enable predictions of their outcomes. Their work is mostly figuring out how to get to the desired result, and how to do so with the most accuracy. Ian Hacking, in his introduction to Kuhn’s <em>Structure of Scientific Revolutions</em> wrote:</p><blockquote><em>Normal science does not aim at novelty but at clearing up the status quo. It tends to discover what it expects to discover. Discovery comes not when something goes right but when something is awry, a novelty that runs counter to what was expected. In short, what appears to be an anomaly.</em></blockquote><h3>Paradigms can change and they do so because we change opinions.</h3><p>A paradigm shift is essentially revoking the accepted axioms. Comparing the paradigms before and after the shifts is not an easy task, and it will be anything but objective. Most of the time it is not feasible to find one paradigm superior to the other, because they tend to differ on fundamental questions. Moreover, the new paradigm will usually leave a lot of questions unanswered (as every paradigm does, especially more the recent ones) and there will be people that won’t adopt it on this basis.</p><p>Kuhn explained that when a new paradigm is adopted, there is always opposition to it. Some researchers will tend to pursue their work just as it was before the shift occurred, and they will remain faithful to their ideas until their very last days. A change will take place nonetheless, as new scientists will go through education learning about the new paradigms and eventually replacing the older generations. This is common characteristic of a lot of changes. Because conservatives tend to offer a counterweight to the overly optimistic innovators, it’s possible to keep the balance between the new and the old.</p><h3>When we change the paradigms, we make the most progress.</h3><p>While most science progresses when paradigms are stable, the <strong>most progress</strong> in science happens when the paradigms are changed. We already know <em>what</em> a paradigm shift is, but let’s talk about why does it happen. Firstly, notice that paradigms are in no way objective, i.e. different groups might clash on particular questions and there is no answer that can be proved superior to others. The emphasis on which questions are important at a given time makes certain sets of axioms more suitable. When people find that their paradigm is unable to explain what is happening in the world, they look for new explanations.</p><p>Problem solving under a paradigm shift is different from problem solving during stable times. Instead of looking to come up with a more or less expected answer in creative ways, scientists relax fundamental assumptions and test which changes can explain the reality better. A new paradigm can become more popular, but never fully accepted, as it is not possible to convince everyone, especially on important topics. Also, there is no clear way to pinpoint <em>when</em> a new set of rules is really accepted, because more often then not, the novel approach needs time before it develops and becomes useful.</p><p>Think about the invention of a neural network. The concept was first proposed in 1943, by two researchers in order to present a simplified mathematical model of neurons. It took 14 years before the first trainable neural network, called the perceptron, was proposed in 1957. After this period, the idea of neural networks lost interest, as von Neumann’s architecture took over the computing scene and popularized a logic based approach. By that time, most people have lost interest in neural networks, and discarded them as a plausible way to arrive at artificial intelligence. Then, with the idea of backpropagation suggested in 1980s neural nets shined again. A crucial moment came in 2012, when AlexNet was presented and went mainstream. In 2024, the Nobel Prize in Physics was awarded to John J. Hopfield and Geoffrey Hinton for their <em>foundational discoveries and inventions that enable machine learning with artificial neural networks</em>. It is unlikely that such an award would be given if backpropagation and AlexNet did not come about. It is also hard to say at which point did neural networks really took off and who in particular was their inventor.</p><h3>New paradigms do not get us close to the truth.</h3><p>Truth exists, perhaps only outside of Plato’s cave and is not accessible to us. Science does not progress <em>towards the truth</em> or <em>away from falsehood</em> in a linear fashion of any sorts. We stumble, make mistakes, embrace ideas which are worse then their predecessors. This happened in the past, and it will happen again. It’s foolish to believe that we can objectively evaluate our current approaches and judge that they are better than previous ones. A true scientist acknowledges that his understanding is limited, yet works on deepening it. The scientific method is not infallible, and there have been times where it lead us astray. Yet, it might just be the best thing we have.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ae0a5f6d015" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Did You Ever Wonder How Elevator Algorithms Work?]]></title>
            <link>https://medium.com/@ignacyrusz/did-you-ever-wonder-how-elevator-algorithms-work-7bf1e0161034?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/7bf1e0161034</guid>
            <category><![CDATA[history]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[monte-carlo-simulation]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Sat, 20 Jun 2026 06:46:56 GMT</pubDate>
            <atom:updated>2026-06-20T06:46:56.041Z</atom:updated>
            <content:encoded><![CDATA[<h3>Story About Lifts and What Do They Have to do With The Hydrogen Bomb.</h3><h3>Press and button and wait.</h3><p>That is what most people do when they encounter a lift, or an elevator if you come from the US. Press the button and when the lift arrives press another button, you are guaranteed to get wherever you are going in a convenient way. Like most people, right? When I moved to a place where I used a lift to get to my floor I started wondering how the algorithms behind the lift work behind the curtains. What does the lift optimize for, the smallest mean or the smallest variance? Does it always go the the ground floor when idle for some time? If so, after how long does it do so? Does it differ during the day and the night? Why would that be?</p><p>If you are anything like me, you came up with similar questions yourself. In this article I show some of my findings about how the intriguing systems behind elevators work, and I will answer the question of what do they have to do with the hydrogen bomb.</p><h3>It’s so simple, they surely just use some optimization algorithm.</h3><p>Yeah, well, they do. But, what optimization algorithm? What is the objective that the algorithm tries to optimize? What is the best metric to optimize after all? After pondering about how these systems might work for some time you might find yourself coming up with different metrics that they might be optimizing for. A few reasonable guesses could be mean waiting time, mean ride time, energy usage, number of people serviced in a unit of time, or perhaps one could care about optimizing some other metric than a mean, like variance or the chance of some extreme event not happening.</p><p>As you can probably tell by now, there are quite a few different things one can optimize for when working on the algorithm of a seemingly simple lift. Actually, there are whole fields that tackle these subjects, in particular look into the queueing theory subfield of operations research if you want to see how did people in the past figure this stuff out. Modern elevators are quite efficient at what they do — transport passengers. If look at them per trip, or possibly by the time their users are exposed to them, they rank among the safest modes of transport, closing in, or surpassing even commercial flights. That being said, there is quite a bit of interesting technical details that go into how lifts function. While doing research for this post I stumbled across some interesting facts about lifts, like their <a href="https://en.wikipedia.org/wiki/Elevator#Special_operating_modes">special operating modes</a>, but in this post I will try to limit my focus on the algorithmic part.</p><h3>Okay, what are the lifts optimizing then?</h3><p>Among the most widely used elevator algorithms is <strong>SCAN</strong>, which makes an elevator continue travel in the same direction (up or down) while making stops to let people off or to pick them up, if they are heading in the same direction. Lifts moving with this algorithm always move to the last floor in a particular direction, and reverse only then. SCAN looks to optimize travel time, i.e. waiting time + ride time. A popular, more efficient version of SCAN is called <strong>LOOK</strong>, which behaves the same as the previous algorithm, with the difference being that it only goes as far as the last requested stop in a particular direction, and not the last possible stop, before reversing to the other direction.</p><p>Both of these algorithms have an important feature, they avoid what is called <em>starvation</em>. Imagine a scenario with a single lift in a building with one hundred floors, optimizing the mean travel time, with majority of the people in the lobby on the ground floor, travelling to floors 1–5, and a single person on the floor 100. If people keep coming to the lobby and pressing the button, the lift that just optimizes the mean travel time wouldn’t go all the way up to pick up that one passenger from floor 100, as long as there would be passengers on the lower floors. That’s called <em>starvation</em>, because one user is <em>starved</em> of the service while others use it repeatedly, by using the SCAN or LOOK algorithms lifts are able to ensure such situations don’t happen. The example of starvation shows that looking to minimize only one metric for cost, is not enough (wink wink my finance friends) because optimizing for low averages doesn’t take into account extreme scenarios, which is not desirable.</p><p>In reality, in buildings which have a lot of floors, there are usually lifts servicing only a selection of them, effectively partitioning the building into smaller parts. For example, lift “A” travels only between the ground floor and floors 1–15, while lift “B” starts at floor 0 and travels between floors 16–30. There can also be high-speed lifts, going rapidly from the ground floor to the upper levels.</p><h3>Right, what if we have multiple lifts in a building?</h3><p>The setup if quite simple if we talk about a single, relatively short, building and a single lift. But when buildings get taller and multiple lifts are operating things start to be a little more interesting. Some skyscrapers make users choose the floor they want to go to before entering the lift itself by choosing a button on a terminal screen in a lobby. The people are then informed, on the same terminal screen, the lift number they should get into to travel to their destination. When they get into that elevator, they can’t change their minds, as their are no buttons inside the lifts, they can only call a new one once they get out. This way, destinations control systems can be employed to optimize traffic in high-rise buildings.</p><p>In such buildings, people wait for their designated car, and do not simply board the next available car. This allows to reduce travel time, enabling elevators to make fewer steps for individual passengers. Despite travel time being reduced, the trip might seem longer because the time one has to wait to get into the lift is increased. To avoid one user requesting multiple lifts at a time and messing up the queue, some systems make everyone use an RFID card which they have to scan when ordering the lift. This way, when a user wants to change their destination of travel by calling the elevator two times, the system doesn’t double count him, but instead cancels the first ride and orders him a second one.</p><h3>What about the hydrogen bomb?</h3><p>Yes, the bomb. But first consider this; as you have already seen operating multiple lifts at a time can get quite complex. If you watched the Nolan’s <em>Oppenheimer</em> you might recall one particularly with a pair of bushy brows. That person was Benny Safdie, playing a person with an even more bushy pair of brows, a Hungarian mathematician and physicist Edvard Teller, who worked on the Manhattan project, focused on building the first nuclear bomb. Teller’s involvement in Project Manhattan was remarkable, but even more remarkable was his involvement in the design of the hydrogen bomb, a bomb thousands times stronger then the nuclear bombs dropped on Hiroshima. There you have the bomb, now, wait a bit and you will understand how it connects to the elevators we have been talking about.</p><p>While <em>Oppenheimer</em> was, in my opinion, great, it failed to portray the second father of the H-Bomb, a Polish mathematician called Stanisław Ulam who was also involved in Project Manhattan. Ulam’s input into the design of the H-bomb was big, and there are debates going over who of the two designers contributed more. Ulam however, is also remembered for another invention he made.</p><figure><img alt="A photo of Stanisław Ulam smoking a cigarette" src="https://cdn-images-1.medium.com/max/1024/0*GrpmdrQlMjF70Ruj" /><figcaption>Stanisław Ulam, from family archives.</figcaption></figure><h3>Tailor-made lift algorithms.</h3><p>Because of the complexity of systems that utilize numerous lifts in buildings with a lot of floors, there is demand for algorithms optimized specifically for that particular buildings, at particular times of operation. One, efficient, way to achieve this is through computer simulation. One of the most widely recognized kinds of computer simulation is the second popular invention of Ulam, namely the Monte Carlo simulation, named after the casino in Monte Carlo that has a lot to do with probabilities and randomness. The Monte Carlo method utilizes randomness to solve deterministic problems. It enables to derive approximate solutions to problems too complex for mathematical analysis, such as the ones within big and complicated building’s lift systems. While the method deserves a post of its own, luckily there have been thousands of people on the internet who wrote about how it works.</p><p>In short, by replicating the lift system of a building within a computer environment, and modeling the flow of people within a building, you can run hundreds of thousands of simulations with different strategies and measure metrics of choice in the process. By following this process you can get closer to picking a tailor-made, optimal solution to implement in your skyscraper, if you ever happen to own one. This is how systems learned that during daytime it might make sense to have lifts idle on the lobby floor, while in the night it might make more sense to keep some of them in the middle of a building or elsewhere.</p><p>There is a lot of interesting tangents that one can go off when trying to optimize different objectives. Most importantly, you do not have to be a real estate owner to think about such issues. If you care about them from a purely intellectual perspective, you will be just fine grappling with different online simulators, like this <a href="https://play.elevatorsaga.com/">elevator programming game</a>, or building your one. If this whole elevator algorithm thing grabbed your attention, expect more of such content in the future here.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7bf1e0161034" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Becoming Dependent in the World of AI]]></title>
            <link>https://medium.com/@ignacyrusz/becoming-dependent-in-the-world-of-ai-3e9b73b42207?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/3e9b73b42207</guid>
            <category><![CDATA[economy]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[studying]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 09:18:02 GMT</pubDate>
            <atom:updated>2026-06-10T09:18:52.691Z</atom:updated>
            <content:encoded><![CDATA[<h3>Becoming Dependent in the World of AI and Why Do we Need a Don Quixote to Save Us.</h3><h3>We were not designed to do boring tasks.</h3><p>Our nature demands and rewards exploration, constantly looking out for new challenges and expanding our understanding. Yet, there are many seemingly boring tasks our society, on the other hand, demands to be done. We are trying to outsource them all to AI, without considering the hidden costs and risks.</p><p>Algorithms can automate and speed up a lot of the boring and non-creative work, completing it faster and more accurately. Obviously, this is a great thing, and has enabled us to dramatically improve the quality of our lives, but we cannot let every task be overtaken by an algorithm, as there is a need for human supervision. More and more companies are buying enterprise versions of the most capable AI models and replacing whole departments of people, who have experience and agency in their job, with less experienced staff equipped with the latest and greatest models. They change labor costs to token costs.</p><p>Meanwhile, individuals get lost, running to increase their productivity in every possible way, trying to meet the expectations put on them by our society. They see AI as a perfect way to multiple their work output, handing over more of the mundane tasks they have to do, and increasingly more of the ambitious and demanding tasks as well.</p><h3>When in Rome, do as the Romans do</h3><p>We live in a society that expects constant growth and gains. There is an expectation that all companies will increase their revenue year by year. That’s how our economy works. The environment demands the perceived increase of productivity, and the companies need to comply by showing the people who lend them money that they are working on it, and they are working on it hard.</p><p>Businesses have went from being seen as innovative and forward looking when implementing AI, to being seen as too conservative and incapable of adapting to the current needs when <em>failing</em> to adopt AI across their business processes. This shift of perspective put a lot of pressure on the directors and managers of all companies to incentivize the use of AI tools to increase the company’s overall (perceived) productivity and as a consequence, the employees need to comply.</p><p>We have seen this trend with technologies in the past, companies rushed to make use of technologies like Cloud or experiment with Big Data. This time it is different, it is different because people have to let go of their understanding of core parts of what they do, in the name of perceived productivity gains, which are expected of them. This is a trade-off because they can’t spend as much time as previously on the same tasks, they can complete them faster with the help of AI assistants.</p><p>This paradigm shift did not stop at affecting businesses only. Students across schools and universities are incentivized to use AI more and more. Teachers themselves use it to grade their pupils’ work. <a href="https://en.wikipedia.org/wiki/Dead_Internet_theory">Dead Internet theory</a>, according to which there are no human interactions on the internet anymore, only bots interacting with each other, seems to be realized in classrooms across the globe.</p><p>Assignments get harder and more time-consuming as universities can’t find proper ways to cope with the rise of new technologies, raising the bar with the expectation of everyone cheating on their tasks. At the same time, students have more trouble understanding the content of the courses, but <a href="https://www.dailycal.org/news/campus/academics/failing-grades-soar-as-professors-see-greater-ai-usage-dwindling-math-skills-in-uc-berkeley/article_16fad0bf-02cb-4b8c-8d88-888ffd9f8608.html">do not realize it until the exams are due</a>. This problem also affects well-known universities.</p><h3>Is there a puppet master?</h3><p>In Cervantes’ <em>Don Quixote</em>, the main character and his trusty squire, Sancho Panza, settle down in an inn to spend a night, when a well-known puppeteer comes in and starts giving a puppet show. The courageous Don Quixote interrupts the show two times before launching a full-scale attack, massacring the puppets with his sword, barely missing the puppet master himself. He did so because he felt that some of the puppet characters needed his help. While Don Quixote admitted his attack was a mistake, regretting it after realizing what happened in reality, he surely did put the puppet show to an end.</p><figure><img alt="Don Quixote demolishing a puppet show with his sword." src="https://cdn-images-1.medium.com/max/940/0*NfOGYcp8ptIi_opY" /><figcaption>Salvador Tusell (fl 1890–1905), Illustration for ‘El ingenioso hidalgo Don Quijote de la Mancha’ (c 1894), watercolour after Gustave Doré, dimensions not known, Fondo Antiguo de la Biblioteca de la Universidad de Sevilla, Seville, Spain. Wikimedia Commons.</figcaption></figure><p>While there might not be a single puppet master in our world today, there surely are powerful actors that have incentives in wanting to implement AI as fast as possible and wherever possible. As time goes on, more and more progress on this aim is being made, and there doesn’t seem to be an actor powerful enough to put this to a timely halt, so that we can consider what is really going on and how we can safely progress into the future. Individuals and whole economies are losing independence in doing work without using AI tools.</p><p>Several problems arise. Obviously, a situation where a few particularly strong companies have essentially a monopoly on the most powerful models, which will be running a big part of the world’s economies, is not ideal, given that they can raise their API prices or equivalently, cut down token limits as they wish. Then, there is also a bunch of concerns about the safety of such wide adoption of these tools, given that we can’t perfectly control their usage and don’t know their exact limitations. These systems can be implemented in a safety-critical environment without people paying too much attention. This year, for example, <a href="https://codewall.ai/blog/how-we-hacked-mckinseys-ai-platform">McKinsey’s internal chatbot got compromised</a>, interestingly enough, by another AI agent leaking secret information about some of the biggest companies in the world.</p><p>Resisting the urge to use AI models gets harder because not only are the environments that surround us demanding it more, but the AI companies are giving ordinary users (or ordinary paying users) access to their more capable models for a period of time, only to lock them behind more powerful paywalls afterwards. This is how Antrophic, the company behind Claude, decided to release their new Fable model. They explicitly say their Pro plan users will have access to it for the 12 days after its release, and they will move it to a more expensive pricing model afterwards. This way of marketing is nothing new, but it is hard not to see the resemblance between such practices and drug dealers offering a sample of some <em>new stuff</em> to get their clients hooked on it.</p><h3>Remember nuclear warheads?</h3><p>With the outsourcing of <em>doing the actual work</em> to AI systems, we are losing important working knowledge and the talent who knows how to do key things. This is not a problem we haven’t seen before. <a href="https://en.wikipedia.org/wiki/Fogbank">Fogbank</a> is a codename of a secret material, critical to craft nuclear warheads, whose production process was lost as the original engineers working on it retired or have deceased, without passing on the tacit knowledge needed to create it. The National Nuclear Security Administration spent 5 years and tens of millions of dollars to reverse engineer the material and how to produce it afterwards. The knowledge on how to create it was not written in any documents, and it could only be passed on by practicing engineers.</p><p>A similar process is prone to happen as we outsource real work across all sectors to AI and have fewer and fewer people actually thinking about it. Not documenting working knowledge properly, we are bound to see it disappear. After some time, we might find ourselves in a similar situation as the US military did in the early 2000s, noticing that we forgot some fundamental process to craft something we need. Only, unlike the US military, we might not have the time or the budget to make up for that loss.</p><h3>Become irreplaceable</h3><p>Knowing how to use AI seems to be an essential skill in today’s workplace, but knowing how to work without AI might be a critical skill in the workplace of the future. This should not be understood in the context of work only, as we have a tendency to take the easy route whenever possible. We should think hard about stuff that matters to us, and as I explained in <a href="https://medium.com/@ignacyrusz/if-you-care-dont-guess-157a3fdb0cde">one of my previous articles</a>, don’t outsource it completely.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3e9b73b42207" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Studies show that the number of friends you have significantly lowers when you turn 25.]]></title>
            <link>https://medium.com/@ignacyrusz/studies-show-that-the-number-of-friends-you-have-significantly-lowers-when-you-turn-25-a33addbdbafd?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/a33addbdbafd</guid>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[friendship]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[change]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Tue, 30 Dec 2025 13:29:28 GMT</pubDate>
            <atom:updated>2025-12-30T13:29:28.281Z</atom:updated>
            <content:encoded><![CDATA[<h4>You can reverse this trend by following a few simple techniques.</h4><p>Make activity friends to maximize your benefit. Join a group that participates in regular activities so meetings are scheduled for you, and you’ll meet new people easily.</p><p>Instead of meeting one-on-one, meet people in groups to maximize exposure. Helps you work around time constraints.</p><p>Host meetings yourself to make it easier for others to join.</p><p>Good things happening to people are mostly the result of conscious effort. Nassim Nicholas Taleb describes in his bestselling book <em>The Black Swan</em> the concept of Domain Dependence. In short, it explains why people fail to generalize concepts from contexts they know to contexts they don’t. People make conscious decisions to exercise and eat healthy because it has a positive effect on longevity and quality of life¹. At the same time, they won’t make such an effort to nurture their social relationships, which have been shown to be a significant factor affecting wellbeing².</p><p>Science shows that at around age 25, the number of social connections people have starts to decline. This trend doesn’t stop as we get older, and is likely to accelerate in today’s age of social media³. To feel good, one needs to <strong>consciously take care of his friends circle</strong>. You have to become the initiator of interactions, otherwise, you will be left to rot. Luckily, there are ways to do this efficiently and reverse the negative trend of losing friends with time.</p><h3>Meeting new people expands your understanding.</h3><p>By being exposed to different perspectives, your vision broadens. You look at arguments from more angles and have a better grasp of them. The most original ideas and solutions come from connecting concepts by far analogies. Bouncing your ideas off others helps you develop them and make better decisions in times of uncertainty. There is a reason why, when the FBI negotiates in critical situations, they have multiple people listening into the line and proposing ideas in real time.</p><p>Stefan Banach, widely considered as one of the most influential mathematicians of the twentieth century, said:</p><blockquote><em>Good mathematicians see analogies between theorems or theories. The very best ones see analogies between analogies.</em></blockquote><p>This is true not only about mathematicians, but about people in general. The more creative you are, the more you can innovate and make the world a better place.</p><p>Another benefit of new connections is that you can discover new ideas or activities to engage in. There are things you would find enjoyable if you were introduced to them in the first place. The problem is that people can live their whole life without ever discovering them. It is not uncommon to meet people who have experience in things you have always wanted to try out. Spending time with them serves as a perfect way to try out these new activities, while having somebody to guide you.</p><h3>Make activity friends to meet them regularly and avoid the awkwardness of scheduling.</h3><p>Activity friends are people you meet while doing some group activities. These include being a member of your local debating club, going for hikes with others, or joining an amateur weekly bouldering session. As people age, they have more duties and less time to meet with others. Scheduling meetings with every person you know would take a lot of time and adjusting, often failing completely. This is not because of the bad intentions of others, it’s simply that they have higher priority commitments. This is where being involved in an activity group fits perfectly.</p><p>Usually, the meetings are scheduled ahead of time at regular intervals. You can expect to join the debating club every Monday evening and see some familiar faces. Certain people might not make it in a particular week, but they will surely join the next session or the one after that. Nevertheless, there will always be friends whom you can talk to. The activity you do is also clear, in this example, debating, so there is no pressure to invent something new every week. The type of hobby matters for the quality of the relationship you can have. Some things, like watching movies, do not require interactions between people. Other commitments involve more contact, like a dancing class, which makes it easier to form bonds. As much as you can try to engage in activities requiring a lot of interactions, while transforming the less interactive ones by saving time to talk with others about what you saw or read together.</p><p>When you feel like it, there is great potential in proposing to grab drinks or some food after a book (or some other) club meeting finishes. This way, people don’t feel the pressure to join, but given they are already there, and you seem like a cool person, given you initiate such things, they will be happy to join. This touches on the next technique that is a must-use.</p><h3>Meet people in groups to maximize exposure and counteract a lack of time.</h3><p>If you can meet a group of 13 people in an hour, you would have to spend 12 additional hours to meet them all one by one. The math here is simple. In order to counteract the natural order of things, you need to optimize your interactions and meet in groups. This is not to say that you can’t meet people individually, because <em>it’s not optimal</em>. You shouldn’t care about optimizing everything Huberman-Lab-style, that’s not realistic. But take into account what you are doing with your and others’ limited time. Because the number of responsibilities a person has increases with age, you need to make the most of the time that is still free. Commitment to meeting in a group is bigger if you know everyone else will show up. You don’t want to make the other person feel FOMO, but because everyone else is showing up, they feel the need to do so as well. Such commitments help sustain relationships through the busiest periods, where you have to focus on family or career.</p><p>Meeting people one-on-one is a good idea to nurture your closest connections. It is also a better setting for talking about more vulnerable topics or discussing serious decisions and fears.</p><h3>Host meetings at your place to decrease friction and make people you invite more likely to show up.</h3><p>While this approach seems contrary to being time-efficient, it really is. By inviting multiple people for an evening of wine and snacks at your place, you do not have to pick and choose who to invite and coordinate timelines. You just tell everybody on what day and time you host something, and tell them to feel free to join. This way, showing up doesn’t require a lot of effort and makes the planning process easier for you. By asking people to do you a small favor, like bringing some ice for the drinks, you invoke the Ben Franklin Effect. It is a term from psychology explaining how people like you more if they do small things for you. This is because their subconscious treats this signal as, <em>I am doing something for this person, therefore I must like him — if I didn’t like him, I would do it!</em></p><h3>But a human can only have 150 meaningful social relationships. Why would you try to have more? It’s not possible.</h3><p>I am not trying to argue that this is not true. Instead, if you can only have 150 friends, you should have the <em>best</em> 150 friends you can have. We naturally lose contact with people over time. This frees up the time to meet new people in return, enabling you to meet and retain only the most interesting relationships.</p><h3>Some math to illustrate the patterns above. Feel free to skip.</h3><p>There are some figures that science has arrived at, trying to describe humans’ relationship over a longer period of time.</p><ol><li>150 People is the maximum number of individuals you can regularly have contact with,</li><li>At 25 years of age, the number of social contacts peaks, and then sharply declines,</li><li>We lose 50% of friends every 7 years,</li><li>This translates to losing about 9–10% of friends every year.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*EjJ1uEMIvm_d8d6F.png" /></figure><p>Clearly, we are fighting an uphill battle here. To stand a chance at winning it, it helps to be aware of the numbers. To help imagine this problem better, I built a simulation using Claude and Vercel, presenting human relationships over time, and how making new friends defeats the natural trends. You can check it out under this <a href="https://friendship-simulator.vercel.app/">link</a>. Below you can see a GIF presenting a few manipulations of the variables.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1008/1*ppfaTwCmzSmEZOj8X1g--A.gif" /></figure><p>To make up for the 9–10% of friends we lose every year, given we start from Dunbar’s sympathy group of 15 people, we should be aiming at meeting and developing a good relationship with at least 1.5 people a year. That is three people in two years. Of course, it is not that straightforward, and forming relations with people is not as simple as dividing and multiplying some numbers. From the simulation and its assumptions, we see that aiming to meet 1.5 new friends per year would require roughly 300 hours of effort per year. That is 7.5 working weeks. It’s a lot of time, but it helps to be aware of the reasons and fulfill these obligations you have to yourself. After all, it’s one of the most important factors for your well-being and quality of life over time.</p><h3>The math is not always the same.</h3><p>While I write about how it seems enough to make 1.5 new real connections a year to have a lot of friends, it doesn’t work like this in every setting. Cultures and environments differ. There is one special scenario, which not only naturally makes it easier to make new friends, it also connects you with people from different cultures and helps form a strong connection for life without needing as much effort.</p><p>Bibliography<br>1. Ford, E. S., Bergmann, M. M., Kröger, H., &amp; Boeing, H. (2013). Healthy living is the best revenge: Findings from the European Prospective Investigation into Cancer and Nutrition–Potsdam study. <em>Archives of Internal Medicine, 169</em>(15), 1355–1362.<br>2. Umberson, D., &amp; Montez, J. K. (2010). Social relationships and health: A flashpoint for health policy. <em>Journal of Health and Social Behavior, 51</em>(Suppl), S54–S66. <a href="https://doi.org/10.1177/0022146510383501"><strong>https://doi.org/10.1177/0022146510383501</strong></a>​<br>3. Bhattacharya, K., Ghosh, A., Monsivais, D., Dunbar, R. I. M., &amp; Kaski, K. (2016). Sex differences in social focus across the life cycle in humans. <em>Royal Society Open Science, 3</em>(4), 160097. <a href="https://doi.org/10.1098/rsos.160097"><strong>https://doi.org/10.1098/rsos.160097</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a33addbdbafd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Do You Feel Miserable By Living in a City. A Simple Fix.]]></title>
            <link>https://medium.com/@ignacyrusz/why-do-you-feel-miserable-by-living-in-a-city-a-simple-fix-a3beef5e25a4?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/a3beef5e25a4</guid>
            <category><![CDATA[nature]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[city-living]]></category>
            <category><![CDATA[hiking]]></category>
            <category><![CDATA[fun]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Sat, 13 Dec 2025 12:46:40 GMT</pubDate>
            <atom:updated>2025-12-13T12:46:40.253Z</atom:updated>
            <content:encoded><![CDATA[<p>Most people live in urban areas. The majority of them rarely spend time outside of them. If they do it, it’s often a long flight to a foreign country for a vacation. People spend a lot of money to have mediocre fun in a city while being pressured to perform and act high-status. The costs stretch beyond the monetary and make people miserable. A way to help yourself and feel better again is to reconnect with nature. Really. And it doesn’t need to be as extreme as you think.</p><h4>The Status Quo</h4><p>Living in a city has its obvious costs and benefits. You get access to a bigger job market, higher wages, and services such as healthcare or education, sometimes unavailable outside of the metropolitan areas. You experience greater availability of public transport, more developed in higher-density areas. There are more people around you, which allows for the creation of niche communities. People will be quick to point out numerous other benefits related to living in a city.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QWowttJphge_5_MIxg9eGg.png" /></figure><p>The costs of accessing these amenities are well known. The rent is expensive, with the average rent as a percentage of income being 27.8% in Germany¹, or 29.8% in the Netherlands². The commute takes a long time, with the average one-way commute time in the European Union being 25 minutes in 2019³. Add to that air and noise pollution, and you start thinking about packing your things and moving out of a city. To sustain a lifestyle like that, you will need a decent wage, and most of it you will spend on basic needs⁴.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B0I5lqi3cKKkks67BfbVSg.png" /></figure><h4>The costs we often overlook</h4><p>What people fail to realize is how much having <em>fun</em> in a city costs. Every time you go out to meet with somebody, you spend a little money here and there. You buy food and drinks or pay for the commute. Going out and seeing all these people dressed nicely, you shop for clothing to impress others. Sport is healthy, and after all, everyone should exercise, so you don’t mind spending 30 euros for one hour at the padel court. These expenses are small, but over a long time horizon, they add up. This behaviour is influenced by your environment. And your environment wants you to pay up.</p><p>While most of these activities cost just a little money, it doesn’t hurt. After all, you can’t feel bad about paying to satisfy your basic needs of meeting people or not being hungry. One often feels like they are saving money by going to a cheaper restaurant or ordering less. The reality is, you are not addressing the root problem.</p><p>Slavoj Žižek critiques the materialistic society by likening capitalism’s functions to perverse religions. People compulsively shop for things companies want them to. Individuals feel accomplished when they finally buy the piece they were looking for. By living and spending our time in cities, we are more likely to take part in the pilgrimages to the shopping malls, purchasing products to appeal to the gods of beauty and status created by capitalism. But people are never fully satisfied with such purchases. When one pick-up is finished, hunting for another one starts immediately. You can’t buy happiness, and yet you try hard to do it.</p><p>This has far-reaching consequences. The psychological problems connected with the disability to satisfy yourself through purchasing new things or services haunt people. They would rather disregard their last purchase as unworthy of solving their problems than grasp the fact that the system itself is ill-designed. Dismissing opportunities to understand what is really happening is a common self-defense mechanism. This is a way to preserve your worldview while staying in your comfort zone.</p><p>The main character of Fyodor Dostoyevsky’s novel <em>Notes from the Underground</em> is presented with an opportunity to understand what his life really is. He is a miserable man, despising all his friends and hating his life. One night, he goes to a brothel where he encounters a prostitute named Liza. He starts giving her a speech about how terrible her life is, how she wastes her youth, beauty, and health. Liza, deeply moved by the speech, is moved. She starts crying. In turn, the man gives Liza his address. He asks her to visit him so they can talk it over and fix what’s broken. The Underground Man is shocked when Liza actually arrives in his poor, neglected apartment. Instead of opening himself and having a genuine conversation with the woman, he acts hostile. Liza was one of the few people who could open the protagonist’s eyes and help him grasp the world around him. She offered genuine affection and willingness to talk with him.</p><p>The Underground Man correctly identified what was wrong with the world of Liza, but he could not confront the reality of his own situation. He was more willing to cut off the only help that life offered him, to keep his worldview intact. People participating in the capitalist society, especially visible in big cities, are like this. They are blind to the ways their lifestyle harms and limits them from day to day. They are too short on time or on willingness to discover what opportunities might be out there.</p><p>This <em>religion</em> of people living in the cities has its necessities. You need money to sustain the lifestyle. Because you work so hard, you need some rest. But resting is expensive, and so is maintaining social relations in a city. What can you do? Simple. Earn more. And how do you earn more? By working harder, you can achieve better results. This is the natural progression of living in a city induced by aspiring to the values shared by the city-dwellers. If you take a step back, you notice this is circular. You want to live better, so you work harder. But then you are worse off and want to live better even more. This can repeat until you can’t cut it anymore.</p><h4>What is the way out?</h4><p>A successful way to tackle your problems and reset is to disconnect from your daily struggle and look at the bigger picture. No binge watching Netflix won’t help you, nor will finding the end of your Instagram reels feed, although I am not sure if the second one is even possible. Keeping your brain busy prevents it from processing problems you might have. In order to think them through, you need exposure to silence. <br>For prolonged periods of time, preferably hours or days. Locking yourself in the basement won’t be of much help, as you will still be in the same cramped state of mind. Changing your environment can help you regain control.</p><p>You have to reconnect with nature. And how do you do it in a city? You get out. While having a walk in your nearby park is a good first step to clarify your thinking and cope with stress, it might not be enough. You will get the full benefits only if you <em>really</em> go out to the wild, cut yourself off from feeling the societal pressure and striving for status, and reconnect with nature. And it doesn’t have to be a fancy vacation trip to Bali. All you sometimes need is to hop on a local train and drive out of the city.</p><h4>But what better place to go is there than my favorite restaurant?</h4><p>Put on your comfortable shoes and find a trail outside of the city you can follow. If you want to feel more secure, use an app like AllTrails or Komoot to discover an interesting area in your neighborhood. A lot of cities have public transport connections that will get you out of the city, often almost for free. If you have a bike, you can just try biking out of the city. Grab some food for the day and pack some water into your bag. You are free to explore and follow whatever path you would like throughout the day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*z_bUOkWLC-ZzZUWK" /><figcaption>Caspar David Friedrich, <a href="https://en.wikipedia.org/wiki/Wanderer_above_the_Sea_of_Fog">Wanderer above the Sea of Fog</a> 1818</figcaption></figure><p>Mountains have been pictured as the perfect place for recovery since ancient times. These enormous geographical formations make us realize how small we really are. They induce a feeling of cosmic insignificance, the comforting realization of how little we really are. It enables us to feel awe at the great wonders of the world. Usually, they are also detached from civilization. The main character of Friedrich Nietzsche’s <em>Thus Spoke Zarathustra</em> disconnected from society to process his thoughts. He wanted to come find enlightenment, and wandered to the mountains to do that. According to the story, he lived there for ten years and expanded his understanding of the self and the philosophy of life. It was in the mountains that he developed a way to understand the world and came up with the idea of Übermensch, the Overman, an ideal set for humans to transcend themselves. Nietzsche writes:</p><blockquote>I love the forest. It is bad to live in cities: there, there are too many of the lustful.⁵</blockquote><p>Lakes and rivers offer a calming atmosphere with the possibility to enjoy a beach or get into the water. Hearing the water flow is similar to sitting by a fire at night. Our species did it for so long that our bodies can’t help but relax when doing so. Trying to grasp the vastness of a sea or an ocean while being at its coast can feel similar to looking at huge mountains. Walking through a forest helps us calm our senses and meditate while moving. In places like these, one should take it slow and do nothing but listen to the environment or go along.</p><p>Nature offers plenty of possibilities to recharge and reset. They come at little to no price and are more or less accessible to most. Yet, when we are too tangled with our daily lives in cities, we tend to forget how easy it really is to get out. What awaits us in the countryside has the power to drastically improve the quality of our lives and enables us to think and feel better.</p><p>The activities available when we get out of the cities are more calming and accessible than we imagine. They offer great opportunities to get some <em>active</em> rest, disconnect from the buzzing city life, and regain energy. One-day trips are great opportunities for bringing friends or new people along and spending time together. There is no need to spend money at fancy restaurants and dress up in the most stylish clothes. When you are out in the wild, you won’t feel judged all the time. The feeling of having to show up in a new outfit every time will slowly disappear.</p><h4>Conclusions</h4><p>The greatest part about this is that you do not need to move to an isolated cabin in the woods to experience the benefits. There is no need to sell all your belongings, post a <em>goodbye all</em> story for your friends to see, and move out. You can do trips and reconnect with nature in small and regular doses. Every weekend, you can dash off to the closest train station and ride away from the city for a hike. Bring a book for the ride or talk to the people you meet along. The unplanned positive surprises are what you will remember the most after all.</p><p>Not every city is positioned and built the same. Taking an hour train from Munich to the Bavarian Alps is a convenient way to start a hike. But this is rather an unusual setting, and many won’t have it this easy. Different regions offer different landscapes to see and explore. While some are more famous than others, this should not be a concern. The goal is not to impress anyone. After all, you are doing this for yourself. The next time you move, you might consider such criteria, but for the time being, find a way out of the city and enjoy it as much as you can.</p><p>1. <a href="https://www.destatis.de/EN/Press/2023/03/PE23_129_12_63.html">https://www.destatis.de/EN/Press/2023/03/PE23_129_12_63.html</a><br>2. <a href="https://www.numbeo.com/cost-of-living/country_result.jsp?country=Netherlands">https://www.numbeo.com/cost-of-living/country_result.jsp?country=Netherlands</a><br>3. <a href="https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Main_place_of_work_and_commuting_time_-_statistics#Commuting_time_for_majority_of_European_workers_is_less_than_30_minutes">https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Main_place_of_work_and_commuting_time_-_statistics#Commuting_time_for_majority_of_European_workers_is_less_than_30_minutes</a><br>4. <a href="https://ec.europa.eu/eurostat/statistics-explained/index.php?title=File:FIG_3_2024.png">https://ec.europa.eu/eurostat/statistics-explained/index.php?title=File:FIG_3_2024.png</a> <br>5. Friedrich Nietzsche, <em>Thus Spoke Zarathustra</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a3beef5e25a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[If You Care, Don’t Guess. How You Can Turn Struggle Into Understanding.]]></title>
            <link>https://medium.com/@ignacyrusz/if-you-care-dont-guess-157a3fdb0cde?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/157a3fdb0cde</guid>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[studying]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Mon, 18 Aug 2025 11:43:59 GMT</pubDate>
            <atom:updated>2025-08-18T13:21:09.740Z</atom:updated>
            <content:encoded><![CDATA[<p>In a world where accessing information is easier then ever, we should try to take a moment to reflect on what we really mean to understand. It is important to take time to comprehend stuff we find important. If there is something you care about and would like to have a legitimate grasp of, you shouldn’t guess. You should be able to explain it clearly.</p><p>This summer, I am doing my first internship. I have been lucky to be assigned real work and projects to work on. Most of the time, I am cleaning data and extending the current codebase by adding some functionalities. It is my first encounter with making things that are going to be <em>pushed to production</em>. This made me think about and appreciate the value of really knowing what you are doing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zH-CatDjUwmcPf24segovA.avif" /></figure><p>When I used to do projects and assignments at university, there really was no incentive to understand all their parts. Most of the time, you would write some code, run it, see if it produces meaningful results, and proceed if it does. If the output didn’t make sense, you would reiterate until it did. Not much thinking needed. Usually, this technique worked just fine, and after two or three attempts, you produced something similar to what you saw on the slides. Sometimes there were parts of the code you didn’t fully understand. But you wouldn’t care much as long as the output looked good. <em>Just let it be for now</em>, was a common saying among some of my friends. We rarely went back to understand those concepts.</p><p>The assignments weren’t usually marked very strictly. There could be multiple reasons for it, including the fact that I am not studying computer science. The result was that you could slack and get alright grades. Because most students optimize to get a good grade, they don’t think too much about actually learning. It is more demanding and rarely rewarded with an appropriate reflection in your grade. Grades are not the best proxy for comprehension. After all, you would often skip understanding, as that wasn’t optimal for getting the best grades you could.</p><p>When I started working, those <em>just let it be</em> moments stopped occurring. Since I am working in risk management in an investment fund, it is really crucial that all things work accordingly to the procedures. There is no room for error. You can’t skip over a certain part of the process only to finish some task earlier. That just doesn’t slide. I often felt a sense of boredom when being asked to go over the same procedure I had already looked into multiple times. But that was just the way it had to be. It had to be checked at each step, and checked carefully.</p><p>What I am trying to say is that there was no longer room for sloppiness. Just letting it be wasn’t an option anymore. I have to be able to vouch for the changes I am making now, and in a sense, I am liable for the mistakes I make. Being an intern, I do not feel the biggest load on my shoulders yet, but this experience gave me a sense of what it can be.</p><p>So I turned to some advice I heard about becoming a good programmer. <strong>Stop guessing. Understand</strong>. This time, I started to appreciate it. When I was writing or reviewing code, I tried to understand the fundamentals and how the program was really structured. Instead of copying and pasting code that printed out bugs, I tried understanding the error messages and referring to the documentation of commands. This was made easier thanks to MATLAB’s extensive docs.</p><p>And it really helped. I believe this is the only way to progress in things such as programming. Nail the fundamentals. Don’t skip over content, but grasp what you are doing. And while asking LLMs about your code is helpful at times, they tend to be <strong>too helpful</strong> and spoon-feed you stuff you would rather find out about yourself. The thing is, if you struggle more to get something, you will remember it better. That is how learning works.</p><p>Now, you see, easy access to information is a double-edged sword. We, cyborgs, with our smartphones always within arm’s reach, can look up information any time. We can easily offload remembering stuff to the internet. We are able to retrieve facts flawlessly at will, with the use of a simple query. This makes our memory way bigger, as we can access information about almost anything. But there is a caveat to that. We stopped remembering and understanding a lot.</p><p>I am not an advocate of struggling for struggle’s sake. I don’t see a point in memorizing the whole cosine table for every degree. Sure, remembering the most useful ones can be useful, but not more than that. I don’t want to cut corners when it comes to the important stuff. Sure, if you have to knock off some stuff from your to-do list, and you want to do it quickly, go ahead. If there is an easier way to get the same job done, why not? But be mindful not to waste an opportunity to learn something useful. It might turn out that knowing how to do <em>that thing</em> you decided to skip earlier will be useful in the future.</p><p>Derek from Veritasium gave a very insightful <a href="https://www.youtube.com/watch?v=0xS68sl2D70">talk</a> on the topic of learning in a world with access to artificial intelligence. In it, he talked about how we need to put some struggle into acquiring information if we mean to hold onto it. And there is no overcoming of this. You need to put yourself out there and test yourself honestly.</p><p>Sure, go ahead and use the newest LLM to ask questions about your code or anything else but watch out not to get an easy and obvious reply which won’t make you learn anything.</p><p>Stay critical when it comes to inbound information. Try to <strong>understand</strong> things instead of <strong>remembering them</strong>. It is an investment that will pay off later on, when you try to actually recall those things. It will make you less prone to making avoidable errors and, in general, able to make higher-quality work. Producing quality results is really needed in the age of ever-present slop.</p><p>Next time you encounter a bug in your code, recall this article and refrain from pasting the error message into your AI assistant of choice before thinking. Ponder about it for a brief moment, and look up the appropriate commands in the documentation if necessary. Figure it out on your own. Trust me, the second time you face the same bug, it will be easier to manage if you ever make the same mistake.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=157a3fdb0cde" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where Does Creativity Come From and How to Foster It: Stories]]></title>
            <link>https://medium.com/@ignacyrusz/where-does-creativity-come-from-and-how-to-foster-it-stories-e06330069be8?source=rss-95a129948399------2</link>
            <guid isPermaLink="false">https://medium.com/p/e06330069be8</guid>
            <category><![CDATA[history]]></category>
            <category><![CDATA[studying]]></category>
            <category><![CDATA[creativity]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[science]]></category>
            <dc:creator><![CDATA[Ignacy Ruszkowski]]></dc:creator>
            <pubDate>Tue, 03 Jun 2025 11:52:55 GMT</pubDate>
            <atom:updated>2025-06-03T11:52:55.112Z</atom:updated>
            <content:encoded><![CDATA[<p>To better grasp what true creativity is, and how it arises, it is a good idea to scrutinize how the most creative individuals came up with their ideas. I am mainly interested in what kind of <em>previous knowledge or inspiration</em> enabled their biggest inventions and what <em>environment</em> was nourishing them during that time.</p><p>My take is that true creativity comes from:<br>1. Deep knowledge<br>2. Vast knowledge<br>3. Space and time to be creative</p><p>I will develop my arguments in the coming paragraphs, by investigating the works and records of prominent figures from the history of science. I see them as particularly interesting cases to scrutinize.</p><h3>How groups and environment influence creativity</h3><p>Stefan Banach is the man behind multiple significant contributions to the field of functional analysis. Some of their key inventions were named by his surname, Banach Spaces, Banach Algebras, and Banach Fixed-Point Theorem. He also collaborated with others and defined the Banach-Tarski paradox, Banach-Steinhaus theorem, and many more. To put it shortly, he was one of the best, if not the best, polish mathematicians and there are several contributions he made that prove this.</p><p>Despite so many scientific findings, his lifestyle was quite unusual. Being a prominent figure in the Lviv School of Mathematics, he developed ideas in collaboration. But he rarely collaborated within closed doors of academia. He enjoyed being outside, with others. “Kawiarnia Szkocka” (direct translation — Scotch Caffe) was the cafe in Lviv that Banach, together with some of the greatest minds of the 20th century, gathered and discussed. In the cafe, the scientist often drank and smoked, with Stefan being the most prominent drinker of them all. Some went as far as calling him an alcoholic.</p><p>His work ethic seemed, to an outsider, questionable at most. He would often be found surrounded by people and noise. But in this environment, he was able to discuss his ideas with other brilliant minds that were around him. His behavior surely seemed eccentric, but I believe there was, within it, hidden an important part of the puzzle. Banach tremendously benefited from the multiple meetings at “Kawiarnia Szkocka” thanks to his ability to crystallize ideas that were prompted during his discussions with fellow mathematicians.</p><figure><img alt="An image of Stefan Banach smoking." src="https://cdn-images-1.medium.com/max/996/1*2IwOw_e5oA71kch8hX7xFQ.jpeg" /><figcaption>Stefan Banach, smoking.</figcaption></figure><p>The tables in the cafeteria contained many answers. They did it quite literally because the mathematicians would often thrust in unexpected bursts of creative thoughts and would note them down as soon as possible, on the table tops by which they were currently seated. Often the next day, around 11 am students would come to the Caffe, to rewrite the problems their masters jotted down the day before. These bursts show a huge amount of inspiration that was gathered during the discussions that took place in the cafe.</p><p>You need to be made aware of the kinds of people that participated in these discussions. Stanisław Ulam was one of the most prominent students of Lviv’s Mathematics school. Later on, he went on to become the father of the Hydrogen Bomb — together with Edward Teller, to whom most of the credit is given today. He was also the inventor of the Monte Carlo method. Stanisław Mazur often accompanied Banach and Ulam at the same table. They used to form the trio seated at the <em>most important table</em>. He was a brilliant mathematician who made numerous contributions to the geometrical methods in functional analysis, a field that was very popular in Lviv at that time. After the Second World War, he held many important academic positions in the recovering Poland.</p><figure><img alt="Stanisław Ulam, with a book" src="https://cdn-images-1.medium.com/max/675/1*AgWW-hwN2WtV_BKWub7p0A.jpeg" /><figcaption>Stanisław Ulam</figcaption></figure><p>The intellectual atmosphere that dominated the noisy and fumy insides of “Kawiarnia Szkocka” was an ideal place for great creative talent to flourish. Arguably, it was the discussions that were constantly ongoing there, that contributed to multiple discoveries, the people which visited that place ever did. It inspired multiple of them to look for and create similar hotspots in their later work. Ulam, in his diaries, noted that the atmosphere that was created in Lviv was only ever matched by the one in Los Alamos, where he was working on the Manhattan Project.</p><h3>Olympia Academy</h3><p><em>Szkocka</em> was not the only example of a group of people coming together, that sparked grand ideas. Many people are aware that Einstein’s most productive time, referred to as his <em>Annus Mirabilis </em>took place when he was employed as a patent clerk in 1905. It was during that time he published 4 papers that went on to change how we view the world. Remarkably, he had a lot of time to hone his thinking about physics while employed there.</p><p>Fewer people are cognizant of another, one could argue that an even more important factor, led the famous <em>genius</em> to such inventions. This fact is that he was a part of what he referred to as <em>Olympia Academy</em>. The group was mainly composed of Einstein, Conrad Habicht who was a mathematician, and Maurice Solovine, a student of philosophy. The trio used to hold wide-ranging discussions tackling the biggest intellectual ideas and works of their time. They discussed and deeply considered the thoughts of David Hume, Baruch Spinoza, John Stuart Mill, and Henri Poincaré to name a few.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x2TnwrtcpZZ45K8WUjVM4w.jpeg" /><figcaption>Olympia Academy members, from the left: Conrad Habicht, Maurice Solovine and Albert Einstein</figcaption></figure><p>What they would do is engage in a lively and long-lasting debate on the fundamental questions from science and philosophy. The topics they discussed, and the atmosphere that was sparked, invited Einstein to refine his arguments more and more while receiving critical feedback. This engagement fostered Einstein’s critical thinking and years down the road he would with pride remember and acknowledge them.</p><p>Places such as the ones above enable individuals to create a deep understanding of fundamental topics. Through continuous discussion and the need for refining one’s arguments, one can develop skills that are crucial to making discoveries.</p><h3>The vastness of knowledge</h3><p>Being a part of an intellectual group surely helps with developing ideas. But the intellectual curiosity doesn’t stop there. There is another feature that prevails among creative individuals. The development of vast knowledge is of tremendous help in doing original work.</p><p>Bertrand Russel was a Welsh philosopher, logician, mathematician, and public intellectual. He received a Nobel Prize in Literature and made several contributions to logic, such as coauthoring “Principia Mathematica”, and other fields. He was also known for his pacifist and anti-war activism. In one of his essays, he explains his reasons for gathering seemingly useless knowledge, from numerous areas.</p><p>Developing a contemplative habit of mind is one of the main causes he argues for in “The Praise of Idleness and Other Essays”. He worries, that his contemporary world, with its tendency to organize thoughts, accepts knowledge only as far as it can be applied and used. It is as if the value of knowledge is measured solely by what we can do with its outputs. He sees and is concerned about, the trajectory to kill craftsmanship and esthetic values.</p><p>While Russell mainly pushes the case of gathering knowledge for one’s enjoyment, I want to emphasize the effects it has on creativity. It opens up plenty of ways that we can relate, seemingly unrelated, concepts from different fields. It enables one to see analogies from nature and get inspired to create something of equivalent use in their domain of interest.</p><p>As previously mentioned Ulam, in his autobiography, quotes Banach as saying:</p><blockquote>Good mathematicians see analogies between theorems or theories, the very best ones see analogies between analogies.</blockquote><p>I believe this holds in many other fields. When you possess such a vast amount of knowledge in different domains, it is not a problem to come up with a link between two, superficially not dependent, ideas. By establishing such a <em>vault</em> of information from which you can draw at will, you develop a mind map that expands quickly. Today this method is common in terms of building a <em>second brain</em>, or a Zettelkasten. Of course, a physical representation of your knowledge bank can foster creativity, but at its root lies the pursuit of a broad understanding of topics.</p><p>Now the other thing that Russell argues for, idleness, forms an ideal point for the last ingredient to mix in to arrive at great creativity.</p><h3>Space and time to be creative</h3><p>While we are <em>idle</em> our subconscious mind often works hard. That is part of the reason the term shower thoughts was coined. These are the realizations that dawn upon you while you are not focused on a specific problem actively. Your mind comes up with ideas or solutions to obstacles that you used to think about intensively, in times when you are no longer doing that.</p><p>It is often only once we give ourselves some downtime, that we come up with great ideas. Daniel Kahneman, in his book “Thinking Fast and Slow”, explains the division he came up with of the human mind. It is a two-system model. The first, System 1, is responsible for quick, automatic operations that take little to no effort and do not feel under our control. It is based on intuitions, associations, and emotions. The second, System 2, is turned on in periods of concentrated attention. It can allocate the brain’s resources to focus on a specific task and handle it.</p><p>System 1 is the one that sparks creativity. It can generate ideas constantly. They often surface as spontaneous thoughts, sometimes feeling like “Aha!” moments. Kahneman points to intuition as the source of such reflections. He says that once an idea is activated, there is a cascade of related ideas rushing through your memory. In Kahneman’s definition, creativity can be described as the ability of associative memory to work exceptionally well. This means linking very remote ideas by unusual connections between concepts.</p><p>Kahneman’s work shows that consciously blocking out time, to simply think, can be tremendously helpful in coming up with original ideas. During this <em>downtime</em> our System 1 can work relentlessly on the creation process of so many useful insights. It should be no wonder that thinkers such as Nietzsche often went on long walks to crystallize their ideas. They found that to be an essential part of their intellectual process, probably because their creative thinking was at its peak during that time.</p><figure><img alt="Edvard Munch’s Friedrich Nietzsche (1906)" src="https://cdn-images-1.medium.com/max/800/0*EaXOJJXtcezOs31R.jpg" /><figcaption>Edvard Munch’s Friedrich Nietzsche (1906)</figcaption></figure><h3>Final thoughts</h3><p>Although the findings that I went over might seem quite cliché at first, huge value can be derived from following a few simple rules. Sticking to them will open up multiple pathways to becoming more creative.</p><p>1. Go deep, into the things that you learn about,<br>2. Discover a broad range of topics, even though they sometimes seem useless,<br>3. And take time to properly develop your ideas, whether it would be through solitary pondering or an enlightening discussion.</p><p>Let’s stay critical when learning new stuff, and dive deeply into it, employing the best of our critical thinking capabilities. Maybe one day we’ll come up with concepts as world-changing as Einstein’s.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e06330069be8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>