<?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 Sebastian Buzdugan on Medium]]></title>
        <description><![CDATA[Stories by Sebastian Buzdugan on Medium]]></description>
        <link>https://medium.com/@sebuzdugan?source=rss-c5c4e6660118------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*F9kHXZwXGEL5VvUwlKqOUA.png</url>
            <title>Stories by Sebastian Buzdugan on Medium</title>
            <link>https://medium.com/@sebuzdugan?source=rss-c5c4e6660118------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:57:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sebuzdugan/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[Agents Do Not Prefer Open Hardware. Reverse Engineering Just Got Cheap.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@sebuzdugan/agents-do-not-prefer-open-hardware-reverse-engineering-just-got-cheap-b617aa84070b?source=rss-c5c4e6660118------2"><img src="https://cdn-images-1.medium.com/max/1200/0*68n8-dk9XErQTVWu.png" width="1200"></a></p><p class="medium-feed-snippet">A $70 pocket synthesizer now has 345 firmware repositories on GitHub, and its manufacturer open-sourced its own firmware under GPLv3 two&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@sebuzdugan/agents-do-not-prefer-open-hardware-reverse-engineering-just-got-cheap-b617aa84070b?source=rss-c5c4e6660118------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sebuzdugan/agents-do-not-prefer-open-hardware-reverse-engineering-just-got-cheap-b617aa84070b?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/b617aa84070b</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 20:55:10 GMT</pubDate>
            <atom:updated>2026-11-16T20:55:27.318Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Perfect Physics Bought 20 Points. Changing the Planner Bought 58.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@sebuzdugan/perfect-physics-bought-20-points-changing-the-planner-bought-58-05e806b35005?source=rss-c5c4e6660118------2"><img src="https://cdn-images-1.medium.com/max/1200/0*VUEXrNOlHi9pDGrL.png" width="1200"></a></p><p class="medium-feed-snippet">Someone ran 600 experiments on a JEPA world model and started by replacing its learned predictor with the simulator&#x2019;s exact physics&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@sebuzdugan/perfect-physics-bought-20-points-changing-the-planner-bought-58-05e806b35005?source=rss-c5c4e6660118------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sebuzdugan/perfect-physics-bought-20-points-changing-the-planner-bought-58-05e806b35005?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/05e806b35005</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 11:51:01 GMT</pubDate>
            <atom:updated>2026-11-14T11:51:01.502Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[I Reproduced the Blackwell GPU Fingerprint on a Laptop in 14 Seconds]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@sebuzdugan/i-reproduced-the-blackwell-gpu-fingerprint-on-a-laptop-in-14-seconds-02ad1e53b607?source=rss-c5c4e6660118------2"><img src="https://cdn-images-1.medium.com/max/1200/0*xes1D-fL5xn2VJ0U.png" width="1200"></a></p><p class="medium-feed-snippet">A report published on November 12 claims six Chinese frontier models were initialized on export-controlled NVIDIA Blackwell hardware, and&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@sebuzdugan/i-reproduced-the-blackwell-gpu-fingerprint-on-a-laptop-in-14-seconds-02ad1e53b607?source=rss-c5c4e6660118------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sebuzdugan/i-reproduced-the-blackwell-gpu-fingerprint-on-a-laptop-in-14-seconds-02ad1e53b607?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/02ad1e53b607</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 15:46:15 GMT</pubDate>
            <atom:updated>2026-11-13T15:46:39.925Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[I Benchmarked CloudMounter Against rclone on 6.5 GB of S3 Data]]></title>
            <link>https://medium.com/@sebuzdugan/i-benchmarked-cloudmounter-against-rclone-on-6-5-gb-of-s3-data-ec033b2a77d1?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/ec033b2a77d1</guid>
            <category><![CDATA[aws-s3]]></category>
            <category><![CDATA[macos]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 12:11:42 GMT</pubDate>
            <atom:updated>2026-11-13T12:11:42.566Z</atom:updated>
            <content:encoded><![CDATA[<p>One app opened a 5 GB model checkpoint from S3 in 3.1 seconds. The same app needed almost fifty minutes to read 3,515 small images from the same bucket.</p><p><em>Disclosure: this is a paid review. Electronic Team, who make CloudMounter, paid for it and saw the numbers after I measured them. Nothing below was changed for them.</em></p><h3>The Problem With Mounting a Bucket</h3><p>If you train models, your data and your checkpoints probably live in S3. Your editor, your notebooks and half your tools want a folder.</p><p>Mounting the bucket in Finder looks like the obvious bridge. It is also where a lot of quiet slowness comes from, because a mount hides every network round trip behind something that looks like a local disk.</p><p>So I tested one of the paid options, <a href="https://cloudmounter.net?utm_source=sebuzdugan">CloudMounter</a>, against the free one most engineers already use, rclone. Same Mac, same network, same bytes, every file checked against its original SHA-256.</p><h3>What CloudMounter Actually Is Under the Hood</h3><p>The website does not say how CloudMounter mounts anything, so I opened the app bundle. The Mac App Store build (4.18) ships no FUSE driver and no kernel extension. It plugs into Apple’s File Provider, the same system behind iCloud Drive.</p><pre>plutil -p &quot;/Applications/CloudMounter.app/Contents/PlugIns/MountProvider.mas.appex/Contents/Info.plist&quot; \<br>  | grep ExtensionPointIdentifier<br># &quot;NSExtensionPointIdentifier&quot; =&gt; &quot;com.apple.fileprovider-nonui&quot;</pre><p>That one design choice explains every number in this article.</p><p>With File Provider, a file is a placeholder until something opens it. The first open downloads it. After that it lives on your disk, and you can pin a folder to keep it offline. So every test has two numbers: the first read and the second.</p><p>It also means the App Store build is fully sandboxed, and your keys go into the macOS keychain instead of a config file in your home folder. The version you download from the CloudMounter website is not sandboxed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9PuBYCdT1YYWACLBxPZBXA.png" /><figcaption>CloudMounter plugs into the same system as iCloud Drive. rclone on macOS runs a local NFS server. That difference explains every number below.</figcaption></figure><h3>The Benchmark Setup</h3><p>One S3 bucket in eu-central-1, and a seeded data set so every tool reads identical bytes:</p><pre># make_dataset.py (excerpt): same bytes every run<br>blob(f&quot;{out}/large/checkpoint_5g.bin&quot;, 5 &lt;&lt; 30, seed=1)<br>blob(f&quot;{out}/large/shard_1g.bin&quot;, 1 &lt;&lt; 30, seed=2)<br>for i in range(10000):  # a small image data set<br>    blob(f&quot;{out}/small10k/img_{i:05d}.bin&quot;, 50 * 1024, seed=1000 + i)</pre><p>On top of that, a tree of 20 folders with 500 files each. For a ceiling, I uploaded everything with the AWS CLI first: about 49 MB/s for big files and 95 small files per second, with its default of 10 concurrent requests.</p><p>For rclone I used the Homebrew build. On macOS it has no FUSE mount, and Homebrew says so at install time, so the mount goes through the built-in NFS client instead:</p><pre>rclone nfsmount cmbench:my-bucket ~/cm-mnt/rclone \<br>  --vfs-cache-mode full            # run 1: defaults<br># run 2: same command plus --use-server-modtime</pre><p>CloudMounter ran on its defaults. Everything here measures reads, because that is the work a mounted bucket does most in ML: loading checkpoints and walking datasets.</p><h3>Big Files: Close to Line Speed, Then Instant</h3><p>A cold read of the 5 GB checkpoint took CloudMounter 2 minutes 18 seconds, about 39 MB/s. rclone did it in under two minutes, about the speed the AWS CLI uploaded at. On a first read, CloudMounter gets you about 80% of the connection.</p><p>The second read is where File Provider pays off.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nn0cWbuNg8R0L7rHpLbr9w.png" /><figcaption>The first read runs close to line speed for both tools. The second read is the whole story: 3.1 seconds from CloudMounter’s local copy.</figcaption></figure><p>CloudMounter read the same 5 GB again in 3.1 seconds. That is the SSD, not the cloud. rclone, going through the NFS layer, still needed 86 seconds with a full cache. Its warm reads topped out around 62 MB/s, barely above its first read.</p><p>If you load the same weights into a notebook ten times a day, that is the difference between seconds and roughly fifteen minutes.</p><h3>Big Folders: The Flag That Changed My Conclusion</h3><p>Next, a folder with 10,000 files. CloudMounter listed it in 141 seconds, with its encryption switched on. Its team tells me listing gets faster with encryption turned off in the connection settings. rclone on defaults took 469 seconds, and a few minutes later it took 426 seconds again.</p><p>I nearly wrote that down as a clean CloudMounter win. Then I read the rclone S3 docs. By default rclone stores its own modification time as object metadata, and reading it “takes an additional HEAD request” per file. On top of that, the directory cache expires after 5 minutes.</p><p>One flag removes the per-file request:</p><pre>--use-server-modtime   # use the S3 upload time, skip one HEAD per object</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZOKTAk1ORBxteyYDdhzK_Q.png" /><figcaption>Same folder, same bucket. One rclone flag takes the cold listing from almost eight minutes to three seconds.</figcaption></figure><p>With it, the same cold listing took 3.3 seconds. So the fair version is this: with zero tuning, CloudMounter browses big folders far faster than rclone. Against a tuned rclone, its first listing is over forty times slower.</p><p>Once a folder is listed, CloudMounter shows it again in 0.02 seconds. Tuned rclone took 2.5.</p><h3>Small Files: Where Every Mount Struggles</h3><p>This is the test that matters most if you train models: reading thousands of small files one by one, the way a data loader walks an image folder.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2yOgDSWhkwRx9GASjrLsAg.png" /><figcaption>Cold reads crawl for every mount because each file is its own round trip. Once cached, CloudMounter serves the same files at over 5,000 per second.</figcaption></figure><p>CloudMounter managed 1.2 files per second. I stopped it after almost fifty minutes and 3,515 files. All 10,000 would have taken over two hours. rclone did 6.2 per second on defaults and 9.7 with the flag.</p><p>Each file is its own round trip to S3, and a mount reads them one after another. The AWS CLI reached 95 files per second on the same files because it runs ten requests at once.</p><p>Once cached, CloudMounter served the same files at 5,240 per second. The cost is entirely in the first read.</p><h3>Reproduce It on Your Own Bucket</h3><p>Mount benchmarks are easy to get wrong, so here is the part that keeps them honest. Every read goes through a checksum, and the cold and warm passes are recorded separately:</p><pre>def sha256_file(p, bufsize=8 &lt;&lt; 20):<br>    h = hashlib.sha256()<br>    with open(p, &quot;rb&quot;) as f:<br>        while chunk := f.read(bufsize):<br>            h.update(chunk)<br>    return h.hexdigest()<br>secs, digest = timed(lambda: sha256_file(mount / &quot;seed/large/checkpoint_5g.bin&quot;))<br>assert digest == manifest[&quot;large/checkpoint_5g.bin&quot;]</pre><p>The small-file test mirrors a data loader: list the folder, then open and hash each file in order. No tool returned a single mismatched byte across any run.</p><p>Three traps I hit along the way. First, a stopwatch on “mount” is meaningless with File Provider: after unmounting, the folder still lists from cache, so the next “mount” looks instant. Second, when you check that a write landed, count objects with length(Contents), because the AWS CLI&#39;s KeyCount stops being useful once a prefix passes 1,000 keys and paginates. Third, run each tool alone on the network. Two mounts downloading at once will both look slow.</p><p>Write down the rclone flags you used. The same tool is 140 times faster or slower depending on one of them.</p><h3>What to Watch For</h3><p>Disk is not free. The site talks about optimizing local storage, and that holds until you open things: after my reads, 6.2 GB sat on my drive. rclone with full caching kept 6.7 GB, so this is how caching works, not a CloudMounter flaw. Unpin folders you are done with.</p><p>Unmount does not make the folder disappear. After unmounting, the folder still listed every cached file.</p><p>There is no Azure Blob Storage connection. rclone and Mountain Duck both offer Azure, so if your data lives there, CloudMounter is out.</p><p>If you script your mounts, CloudMounter speaks AppleScript:</p><pre>osascript -e &#39;tell application &quot;CloudMounter&quot; to mount (first connection whose name is &quot;Bench&quot;)&#39;</pre><h3>Final Thoughts</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3V7LCvEMXQMhmxV6rvGZUQ.png" /><figcaption>Every read test in one table. CloudMounter wins every second read, a tuned rclone wins every first read.</figcaption></figure><p>For a training job, do not point a data loader at any mounted bucket and hope. Sync the data set down once, or pin the folder in CloudMounter well before the job starts, so the slow first read happens while you sleep.</p><p>For everything else, the choice is about who does the tuning. If you want S3, Google Drive and an SFTP server in Finder with no config files, CloudMounter is the easiest thing I tested, and its cache makes everything after the first open feel local. If you script everything and stream big data sets, a tuned rclone or a sync step still wins.</p><p>Either way, measure the first read and the second read separately. Most mount benchmarks only report one of them.</p><h3>Resources &amp; References</h3><ul><li><a href="https://cloudmounter.net?utm_source=sebuzdugan">CloudMounter</a></li><li><a href="https://rclone.org/s3/">rclone S3 backend docs: modification times and</a> the use-server-modtime flag</li><li><a href="https://rclone.org/commands/rclone_mount/">rclone mount docs: VFS cache modes and</a> the dir-cache-time flag</li><li><a href="https://rclone.org/commands/rclone_nfsmount/">rclone nfsmount docs</a></li><li><a href="https://docs.aws.amazon.com/cli/latest/topic/s3-config.html">AWS CLI S3 configuration: max_concurrent_requests</a></li></ul><h3>Stay in Touch</h3><p>Short takes and discussions on X → <a href="https://x.com/sebuzdugan">https://x.com/sebuzdugan</a></p><p>Practical AI / ML videos on YouTube → <a href="https://www.youtube.com/@sebuzdugan/">https://www.youtube.com/@sebuzdugan/</a></p><p>Partnerships &amp; collabs → <a href="mailto:sebuzdugan@gmail.com">sebuzdugan@gmail.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ec033b2a77d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[One Sentence Took Claude’s Refactor Rate From Under 10% to Half]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@sebuzdugan/one-sentence-took-claudes-refactor-rate-from-under-10-to-half-63dbc0e7c8c7?source=rss-c5c4e6660118------2"><img src="https://cdn-images-1.medium.com/max/1200/0*5gKLQ_C0gubWouWG.png" width="1200"></a></p><p class="medium-feed-snippet">Anthropic measured a single line of system prompt that moves Claude Fable 5.2 from merging duplicate code in under one try in ten to&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@sebuzdugan/one-sentence-took-claudes-refactor-rate-from-under-10-to-half-63dbc0e7c8c7?source=rss-c5c4e6660118------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sebuzdugan/one-sentence-took-claudes-refactor-rate-from-under-10-to-half-63dbc0e7c8c7?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/63dbc0e7c8c7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 19:09:46 GMT</pubDate>
            <atom:updated>2026-11-10T19:10:20.397Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[100 parallel Harbor rollouts on Tensorlake: 47 rollouts/minute at 37¢ per batch]]></title>
            <link>https://medium.com/data-science-collective/100-parallel-harbor-rollouts-on-tensorlake-47-rollouts-minute-at-37-per-batch-46c4f6aefb19?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/46c4f6aefb19</guid>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 13:54:22 GMT</pubDate>
            <atom:updated>2026-11-10T13:54:22.391Z</atom:updated>
            <content:encoded><![CDATA[<h4>From 10 to 100 concurrent sandboxes: 7.9x the throughput, 300 of 300 passed.</h4><p><em>Sponsored by </em><a href="https://tensorlake.ai/?utm_source=medium&amp;utm_medium=sponsored_content&amp;utm_campaign=sebastian_nov2026"><em>Tensorlake</em></a><em>. They paid for this article and reviewed the draft; the numbers below are from my own runs on 5 and 6 Nov 2026.</em></p><p>Harbor, the framework from the Terminal-Bench creators, runs 4 coding-agent rollouts at a time by default, and the Terminal-Bench 2 README example uses the same 4 on local Docker. That is a config default, not a technical limit: Harbor can run far more, and its docs show 100 concurrent rollouts on other cloud sandboxes. On local Docker, though, raising it runs you into your own machine. Every RL step then waits on the slowest rollout in the batch.</p><p>So I ran Harbor on Tensorlake sandboxes at 100 concurrent, same command, two flags changed. A hundred rollouts finished in about two minutes. Scaling from 10 to 100 concurrent rollouts increased throughput by 7.9x. At 100 concurrency, median rollout latency was 110.6 seconds, 23% higher than running a single rollout. Here is where the seconds and the cents went.</p><p>I measured it with Harbor on Tensorlake sandboxes. The answer in one table:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RmRzciHWJisztZ5wDVrAxg.png" /></figure><p>Throughput is rollouts divided by batch wall time. Wall times, median of 3 runs: 96.4 s at 1, 100.1 s at 10, 127 s at 100 (runs of 166, 127 and 118 s).</p><p>The four numbers to remember: 7.9x throughput from 10 to 100 concurrent (79% of linear); +23% median rollout latency at 100 concurrent versus a single rollout; 300 of 300 rollouts passed at 100 concurrent; 37¢ sandbox cost for 100 rollouts, from Tensorlake’s usage API.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jpwvnKgBM31en6M4rmSPsA.png" /></figure><p><em>Figure 1. Throughput at 1, 10 and 100 concurrent rollouts against the linear ideal. From 10 to 100 it grows 7.9x, 79% of linear.</em></p><h3>What I actually ran</h3><p>Everything uses Harbor’s oracle agent. The oracle replays the task’s reference solution, so there is no language model in the loop. That is on purpose: it isolates the part Tensorlake is responsible for, the sandbox, from the part that varies with the model. It also means these are infrastructure numbers, not agent pass rates. Treat 300 of 300 as “the plumbing did not drop anything”, not “the agent is good”.</p><p>Setup, in a virtual environment (the install pulls 130+ packages and took 157 s for me):</p><pre>pip install &quot;harbor[tensorlake]&quot;<br>export TENSORLAKE_API_KEY=...</pre><p>The rollout, one Terminal-Bench 2.1 task whose image Tensorlake has already published, so there is no build step:</p><pre>harbor run --env tensorlake \<br>  --dataset terminal-bench/terminal-bench-2-1 \<br>  --include-task-name terminal-bench/pytorch-model-cli \<br>  --agent oracle -k 100 -n 100</pre><p>-k is how many attempts, -n is how many run at once. Going from 10 to 100 is changing two numbers. Nothing else in the task or the evaluator changes. Harbor works on native Windows too, which is where I ran it.</p><p><strong>One requirement:</strong> 100 concurrent sandboxes needs a credit-based (usage) plan, which has a limit of 100 concurrent sandboxes. The free tier has lower concurrency limits, so you cannot run this exact command on it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Mnf9ASK1xjYmQQAdxxcbQ.png" /></figure><p><em>Figure 2. The only change is where the rollouts run. Harbor’s default is 4 trials at once; each Tensorlake sandbox takes one rollout.</em></p><h3>Scaling efficiency: what 100 at once costs you</h3><p>From 10 to 100 concurrent, throughput went up 7.9x, which is 79% of the 10x you would get if nothing slowed down. The price of the missing 21% is latency: at 100 concurrency the median rollout took 110.6 s, 23% more than the 89.9 s of a single rollout run alone (the 10-concurrent median was about 90 to 93 s). More than half of that gap is in the verifier (40.7 s alone, 52.7 s at 100). Setup moved a little (8.2 s to 8.6 s) and the agent step went from 37.0 s to 42.0 s.</p><p>Reliability held: 300 of 300 rollouts passed at 100 concurrent, and 334 of 334 across all my oracle runs.</p><p>The tail is the part that varies. The slowest rollout in each batch set that batch’s wall time: 163.6 s, 127.2 s and 117.5 s across the three runs. So at 100 concurrency, plan around the slowest rollout, not the median. If your trainer waits for the whole batch, the tail is your step time. If it can use rollouts as they arrive, you get most of the benefit sooner.</p><h3>Where the 90 seconds go</h3><p>Before scaling anything, I split one rollout into its phases (median of 3 runs): sandbox setup 8.2 s (9%), agent step 37.0 s (41%), verifier 40.7 s (45%).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oQ_VExEUU-mQm7z5uOkb1w.png" /></figure><p><em>Figure 3. The same rollout alone and at 100 concurrent. The verifier grows the most (40.7 s to 52.7 s).</em></p><p>Two things stand out. The sandbox is under a tenth of the rollout, so “sandbox startup” is not where RL time goes. And the verifier takes as long as the agent: scoring is half the job, so a rollout holds a full machine for the whole minute and a half, not just for the agent’s turn.</p><p>For the API call itself, Sandbox.create() took a median of 1.27 s from my laptop over 24 creates. Tensorlake’s docs say sandboxes boot in hundreds of milliseconds; as the team explained to me, that figure is the OS boot after creation, not the API round trip, so expect about a second and a quarter from your code to a sandbox you can use, and about 8 s in Harbor until the task container is ready.</p><h3>What it costs</h3><p>This number is measured, not guessed. Tensorlake’s usage page reports the cost of each sandbox, so I pulled the usage export for my account. It lists 334 sandboxes that ran my oracle rollouts, each sized 2 CPU, 4 GB RAM and 30 GB disk (the task asks for less, but Harbor’s Tensorlake environment raises every sandbox to that minimum). Tensorlake charged $1.23 for all of them: about 0.37 cents per rollout, about 37 cents per 100.</p><p>Billing is per second at 7 cents per core-hour, 1.5 cents per GB-hour of RAM and 0.02 cents per GB-hour of disk. CPU is billed on active execution with a floor of 20% of the allocated vCPU, and my rollouts averaged about 0.85 busy vCPU out of 2. That is why the bill lands between the two bounds I had worked out from the price list alone: 28 cents at the floor, 61 cents with both cores busy the whole time. Of the 37 cents, about 47% is CPU, 48% is RAM and 5% is disk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oevnn8ZbqURjHHGYIZuxHw.png" /></figure><p><em>Figure 4. Cost of 100 rollouts: the measured figure sits between the CPU floor and ceiling.</em></p><p>Per rollout: about 0.37 cents. For 100 rollouts: about 37 cents. For 1,000 rollouts (extrapolated): about $3.70.</p><p>Two caveats. The export is Tensorlake’s current-price estimate, not a finalized invoice, and it rounds each sandbox to a whole cent, so I priced the usage quantities directly: that gives $1.19 against the $1.23 total. And it covers every rollout in my runs, not only the 100-way ones; the 300 longest-running sandboxes average 0.36 cents each, so the figure barely moves. Because you pay per second, the slower rollouts at 100 concurrent cost a little more each than a rollout run alone.</p><h3>What this means for 1,000 rollouts</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*baY6WNmJ9idAZTX6Swh6rg.png" /></figure><p><em>Figure 5. The scorecard: what I measured at each level.</em></p><p>Here is the comparison that matters, for 1,000 rollouts. The rows for 4, 8 and 30 are arithmetic from the rollout time I measured (about 95 s), not runs on local hardware, and a local Docker rollout may take longer or shorter; the 100 row extrapolates my 100-way result to ten waves. I have not tested either.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qqSDCsiMTla4pmsAalxPSQ.png" /></figure><p>That is about 20x faster than the default, 10x faster than 8 concurrent and 2.5x faster than 30, with no machine to size, no images to cache and no cleanup to own. At the measured 0.37 cents each, the sandbox time for the 1,000 would be roughly $3.70. Run it yourself before you plan around it: the tail may behave differently at ten waves than at one.</p><h3>Two things that got in the way</h3><p><strong>Large output gets an extra newline.</strong> A sandbox printing 40,000 characters returned 40,002; 100,000 returned 100,006. That is one added newline every 16,384 characters. If you print JSON over 16 KB from a sandbox and parse it, it fails with “Invalid control character”. Do not strip the newlines: that is not a safe fix, because it also breaks any JSON that contains real newlines. Write the result to a file inside the sandbox and read the file instead. Tensorlake says it is working on the root cause.</p><p><strong>The Harbor page has no RL guide yet.</strong> It says rollouts are coming in follow-up guides. What you do above, -k and -n, works today; the RL examples live on a separate page, “Reproducible Environments for RL Rollouts”.</p><h3>What I did not test</h3><p>No language model agent, so no pass rate and no token cost. No rival sandbox on the same task. No local-Docker runs to confirm the 4, 8 and 30 concurrency estimates. No snapshots, no network-policy phases, no Docker Compose tasks. The cost is Tensorlake’s current-price estimate from the usage export, not a finalized invoice. Scale AI’s AgentEnv is out of this piece: it did not run on native Windows, and Tensorlake pointed out that is an AgentEnv limitation.</p><h3>Try it</h3><p>You need a usage-based plan for the 100 concurrent sandboxes. Start here: <a href="https://tensorlake.ai/?utm_source=medium&amp;utm_medium=sponsored_content&amp;utm_campaign=sebastian_nov2026">tensorlake.ai</a>, and the <a href="https://docs.tensorlake.ai/?utm_source=medium&amp;utm_medium=sponsored_content&amp;utm_campaign=sebastian_nov2026">Tensorlake documentation</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=46c4f6aefb19" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science-collective/100-parallel-harbor-rollouts-on-tensorlake-47-rollouts-minute-at-37-per-batch-46c4f6aefb19">100 parallel Harbor rollouts on Tensorlake: 47 rollouts/minute at 37¢ per batch</a> was originally published in <a href="https://medium.com/data-science-collective">Data Science Collective</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ForwardAgent Defaults to No. The Interesting Bug Is What Happens When You Say Yes.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@sebuzdugan/forwardagent-defaults-to-no-the-interesting-bug-is-what-happens-when-you-say-yes-f075441b9aca?source=rss-c5c4e6660118------2"><img src="https://cdn-images-1.medium.com/max/1200/0*TFlGGkVldWXCxCY7.png" width="1200"></a></p><p class="medium-feed-snippet">Click here to read this article for free</p><p class="medium-feed-link"><a href="https://medium.com/@sebuzdugan/forwardagent-defaults-to-no-the-interesting-bug-is-what-happens-when-you-say-yes-f075441b9aca?source=rss-c5c4e6660118------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sebuzdugan/forwardagent-defaults-to-no-the-interesting-bug-is-what-happens-when-you-say-yes-f075441b9aca?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/f075441b9aca</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 19:16:35 GMT</pubDate>
            <atom:updated>2026-11-09T19:17:45.158Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Unplugging the Laptop Made Inference Four Times Faster.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@sebuzdugan/unplugging-the-laptop-made-inference-four-times-faster-393d52e712f4?source=rss-c5c4e6660118------2"><img src="https://cdn-images-1.medium.com/max/1280/0*dTOOPZWDk0ziaBQo.png" width="1280"></a></p><p class="medium-feed-snippet">On an RTX Spark laptop running native Linux, GPT-OSS-120B generates about 15 tokens per second plugged into the wall. Pull the power cable&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@sebuzdugan/unplugging-the-laptop-made-inference-four-times-faster-393d52e712f4?source=rss-c5c4e6660118------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sebuzdugan/unplugging-the-laptop-made-inference-four-times-faster-393d52e712f4?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/393d52e712f4</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 13:26:01 GMT</pubDate>
            <atom:updated>2026-11-07T13:26:01.649Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Same Code, Framed as Safety Work, Got 40% Less Done.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">OpenAI ran a paired evaluation on GPT-6 Astra. Two codebases, identical except for names, comments and documentation. The same coding task&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@sebuzdugan/the-same-code-framed-as-safety-work-got-40-less-done-31c3b965d947?source=rss-c5c4e6660118------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sebuzdugan/the-same-code-framed-as-safety-work-got-40-less-done-31c3b965d947?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/31c3b965d947</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 16:16:48 GMT</pubDate>
            <atom:updated>2026-11-06T16:17:05.993Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Designkit Free Background Remover Review: A Hands-On Test for Product Photos]]></title>
            <link>https://medium.com/@sebuzdugan/designkit-free-background-remover-review-a-hands-on-test-for-product-photos-aa1b0dc4c30e?source=rss-c5c4e6660118------2</link>
            <guid isPermaLink="false">https://medium.com/p/aa1b0dc4c30e</guid>
            <category><![CDATA[product-photography]]></category>
            <category><![CDATA[ecommerce]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sebastian Buzdugan]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 12:11:10 GMT</pubDate>
            <atom:updated>2026-11-06T12:11:10.890Z</atom:updated>
            <content:encoded><![CDATA[<p>Cutting a product out of its background is the most boring part of building a listing. It is also the part that decides whether your store looks professional or homemade</p><p><em>Disclosure: this is a sponsored review. Designkit paid for this article. The testing, the numbers and the opinions are my own.</em></p><h3>Why I tested this one</h3><p>If you sell on Shopify, Amazon, Etsy or eBay, you know the routine. Shoot the product, cut it out, drop it on white, resize, upload. Repeat for every variant.</p><p>Most free tools handle the cutout part fine on easy photos. They fall apart on glass, fur, jewellery and anything with a busy background.</p><p>So I put the <a href="https://www.designkit.com/background-remover">Designkit free background remover</a> through a proper test. Not three hand-picked hero shots, but 15 product photos ranging from boring to genuinely hard.</p><figure><img alt="The Designkit background remover upload page" src="https://cdn-images-1.medium.com/proxy/1*21_ZFbybTfNKvKcuiBMRtA.jpeg" /><figcaption>The upload page. No account needed to start.</figcaption></figure><h3>What you get for free</h3><p>Here is the free tier as it works today, confirmed in the live product.</p><p>You can upload a photo and preview the cutout <strong>without signing in</strong>. No email, no trial, no card.</p><p>To export the full-size image, you sign in with a free account. That gives you <strong>up to 100 free background removals per day</strong>. I checked the counter after exporting: it showed 100 free per day, with 99 left after my first export. My account’s 20 regular credits stayed untouched.</p><p>Uploads accept JPG, PNG and WebP. The file picker also takes HEIC, which is handy if you shoot products on an iPhone.</p><figure><img alt="The download menu with Preview and Full Resolution options" src="https://cdn-images-1.medium.com/proxy/1*DbPcnqMhPkQkF3qgS-iOVA.jpeg" /><figcaption>After the cutout, you pick a quick preview or the full-resolution export.</figcaption></figure><h3>My test setup</h3><p>I used 15 product photos from Unsplash, each 1200 pixels wide. Credits are at the end of the article.</p><p>The easy group was a mug, a backpack, headphones, a handbag, a watch, a camera and a sneaker. The hard group was the reason for the test:</p><ul><li>A clear glass bottle on a pale background.</li><li>A glass of wine with the pour still in mid-air.</li><li>A fluffy teddy bear, for fur edges.</li><li>A pearl necklace inside its box.</li><li>A cosmetics flat lay and a shelf full of potted plants.</li></ul><p>I ran every photo through the single-image flow twice, on two different days, and timed each run from the browser’s network log.</p><h3>The workflow, step by step</h3><p>The flow is three steps and nothing hides behind menus.</p><p>Upload the photo. Designkit detects the subject on its own, there is nothing to click or brush. A few seconds later you see the result over a transparency grid.</p><p>From the same screen you can swap the background to white, black, grey or any custom colour. That covers the classic marketplace white-background shot without opening another tool. For that specific job Designkit also has a dedicated <a href="https://www.designkit.com/background-remover/white-background">white background maker</a>.</p><p>Then hit Download and choose the preview or the full-resolution file.</p><figure><img alt="The sneaker before and after background removal" src="https://cdn-images-1.medium.com/proxy/1*KaZgassiWqkno37bY5LaPQ.jpeg" /><figcaption>A white sneaker on white fabric. Low contrast, clean result.</figcaption></figure><h3>How fast it actually is</h3><p>Designkit’s own tests put a single image at under 3 seconds, so I measured it myself instead of repeating it.</p><p>Across 30 runs, the time from upload to finished cutout was <strong>2.3 to 4.3 seconds</strong>, with an average of about <strong>3 seconds</strong>.</p><p>The AI step on its own, the part where the model actually removes the background, took between 0.9 and 2.4 seconds. The rest is uploading the photo and loading the result.</p><p>In practice it feels instant. You will spend more time picking the next file than waiting for the cutout.</p><h3>The results on hard photos</h3><p>This is where I expected problems. Mostly, there weren’t any.</p><h4>Glass that stays see-through</h4><p>Glass is the classic failure case. Most removers either cut it out as a solid grey shape or eat half of it.</p><p>Designkit kept the bottle genuinely transparent. Put it on a coloured background and the colour shows through the glass, the way it would in a real photo.</p><figure><img alt="A clear glass bottle before and after, placed on blue" src="https://cdn-images-1.medium.com/proxy/1*Sr0Lg0kPDjgeIE5jKUS01g.jpeg" /><figcaption>The blue shows through the glass, the cork and the outline stay sharp.</figcaption></figure><p>The wine shot was harder still. The thin stream of wine was kept, the glass stayed transparent, and the bottle and pouring hand came along as part of the subject.</p><figure><img alt="Wine being poured into a glass, before and after" src="https://cdn-images-1.medium.com/proxy/1*UFurgCCzcApO2eCVO0vDmQ.jpeg" /><figcaption>The pour stream survives the cutout.</figcaption></figure><h4>Fur and soft edges</h4><p>The teddy bear is a good stand-in for plush toys, knitwear or anything fuzzy. The edge kept the individual tufts instead of cutting a smooth, fake-looking outline.</p><figure><img alt="A teddy bear before and after on a white background" src="https://cdn-images-1.medium.com/proxy/1*g2DnmS-dUFWZrVM_OZzgDg.jpeg" /><figcaption>Fur edges stay fluffy rather than turning into a hard cut.</figcaption></figure><h4>The full set</h4><p>Here are all 15 cutouts together. Every photo processed on the first try, with no errors and no manual fixes.</p><figure><img alt="All 15 product cutouts on a transparency grid" src="https://cdn-images-1.medium.com/proxy/1*rIjJmBi7bAn75J_QMetvNw.jpeg" /><figcaption>15 photos, 15 usable cutouts.</figcaption></figure><p>On solid products like the watch, backpack and headphones, the result is ready for a listing straight away.</p><figure><img alt="A watch, backpack and headphones on a clean white background" src="https://cdn-images-1.medium.com/proxy/1*6g9VA_oRf9dr8T8m0TPrHw.jpeg" /><figcaption>Straight from the cutout to a white listing background.</figcaption></figure><p>The full-resolution export came out at the same size as my input, as a transparent PNG with no watermark.</p><h3>Limitations worth knowing</h3><p>No tool is perfect, and this is the one thing I would want to know before relying on it.</p><p><strong>It decides what the subject is.</strong> The pearl necklace came out together with its box. On the plant shelf, it kept every plant and even the hanging light bulbs. Both are reasonable guesses, but there is no way to tell the free tool “only the necklace”. If your photo has several objects, crop it first.</p><figure><img alt="The pearl necklace cut out together with its box" src="https://cdn-images-1.medium.com/proxy/1*2V2yOVHDuiRrL6-p_IcIGQ.jpeg" /><figcaption>It treated the necklace and the box as one subject.</figcaption></figure><h3>Who it is for</h3><p>If you sell physical products and edit your own photos, this fits neatly into the routine. It is fast enough that you will not put the editing off until later, and 100 free exports a day covers most small shops.</p><p>It is also a good fit for content creators who need clean cutouts for thumbnails, and for small businesses without a designer.</p><p>If you need the cutout in a styled scene afterwards, the <a href="https://www.designkit.com/ai-background-generator">AI background generator</a> picks up where the remover stops.</p><h3>Final Thoughts</h3><p>I went in expecting the usual: fine on a mug, broken on glass. Instead, the hard photos were the most impressive part of the test.</p><p>Clear glass stayed clear, fur stayed soft, and every one of my 15 photos produced a usable cutout in about 3 seconds.</p><p>The one real habit to build is cropping busy photos so the tool picks the subject you want.</p><p>If you want to see how it handles your catalogue, <a href="https://www.designkit.com/background-remover">try Designkit’s free background remover with your own product photos</a> and see how it fits your listing routine.</p><h4>Resources &amp; References</h4><ul><li><a href="https://www.designkit.com/background-remover">Designkit Free Background Remover</a></li><li><a href="https://www.designkit.com/background-remover/white-background">Designkit White Background Maker</a></li><li><a href="https://www.designkit.com/ai-background-generator">Designkit AI Background Generator</a></li><li><a href="https://www.designkit.com/">Designkit home</a></li><li>Test photos from <a href="https://unsplash.com">Unsplash</a> by Jeff Tumale, Laura Chouette, Bobby Donald, VVS, Luke Peterson, Bec R., mostafa mahmoudi, Ceyda Çiftci, Kiran CK, pmv chamara, Carter Baran, Kevin Kelly, Wiser by the Mile, Pedram Farjam and Cornelia Ng.</li></ul><h4>Stay in Touch</h4><p>Short takes and discussions on X</p><p>→ <a href="https://x.com/sebuzdugan">https://x.com/sebuzdugan</a></p><p>Practical AI / ML videos on YouTube</p><p>→ <a href="https://www.youtube.com/@sebuzdugan/">https://www.youtube.com/@sebuzdugan/</a></p><p>Partnerships &amp; collabs</p><p>→ <a href="mailto:sebuzdugan@gmail.com">sebuzdugan@gmail.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa1b0dc4c30e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>