<?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 Tihomir Jauk on Medium]]></title>
        <description><![CDATA[Stories by Tihomir Jauk on Medium]]></description>
        <link>https://medium.com/@tihomirjauk?source=rss-d06d4e5831c0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*454rAg4rNMj3gvLr.jpg</url>
            <title>Stories by Tihomir Jauk on Medium</title>
            <link>https://medium.com/@tihomirjauk?source=rss-d06d4e5831c0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 21:59:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@tihomirjauk/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[I Designed a Hologram Without Touching a Laser]]></title>
            <link>https://medium.com/@tihomirjauk/i-designed-a-hologram-without-touching-a-laser-a365df88b538?source=rss-d06d4e5831c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/a365df88b538</guid>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[hologram]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Tihomir Jauk]]></dc:creator>
            <pubDate>Wed, 22 Jul 2026 20:38:08 GMT</pubDate>
            <atom:updated>2026-07-22T20:38:08.728Z</atom:updated>
            <content:encoded><![CDATA[<h3>Forty simulations, one AI agent loop, and a six-figure mistake hiding in the bill of materials</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Laye8Lx17SV5Rk5UoGk0xQ.jpeg" /><figcaption><em>Artist’s impression — AI-generated, not a photograph. No such device has been built. Every other image in this article is real output from the simulations</em></figcaption></figure><p>Picture a glowing image hanging in the middle of a room. No screen, no goggles, no glass box: just light, floating in open air. You can walk around it, see it from any side, and pass your hand straight through it.</p><p>That’s the thing everyone means when they say <em>hologram</em>. Dennis Gabor worked out the real physics of holography in 1948 and collected a Nobel Prize for it. Star Trek’s holodeck made the rest of us want one badly. Nearly eighty years on, it is still stubbornly absent from living rooms.</p><p>I designed one in about a week, and I never turned on a laser or bought an optical bench. The entire device — its physics, its optics, its control loops, its bill of materials — was computed on a workstation by a stack of open-source physics simulators driven by an AI agent.</p><p>The result is a 49-page design study (link at the end). It contains one finding I think matters commercially, one methodological argument I think matters more, and one result I had to throw away. I’ll take them in that order, starting with the machine itself.</p><h3>Two beams that are harmless until they meet</h3><p>Fire a laser beam into a room and focus it hard enough, and at the focal point the light tears electrons off the air molecules. The air stops being a gas and briefly becomes a plasma: a microscopic spark, about half the width of a human hair, glowing for a few tens of nanoseconds before it fades.</p><p>That spark is a <strong>voxel</strong>, a pixel with a real address in three-dimensional space. It isn’t projected onto anything; it is lit-up air.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p8yRjiusXSSyrLvcM4vQ9Q.png" /><figcaption><em>What a single frame actually looks like: a computed voxel image, front and side view. Every dot is one laser flash. The result is line-art hanging in air; a solid glowing block would need far more power.</em></figcaption></figure><p>Move the spark fast enough — around ten thousand of them, sixty times a second — and your eye stops resolving individual dots and starts seeing a picture, the same way a waved sparkler draws shapes in the dark. A Japanese group, Ochiai and colleagues, demonstrated exactly this in 2015 with a system they called <em>Fairy Lights in Femtoseconds</em>. Touchable, glowing, floating imagery. It works.</p><p>The problem is the beam. A single beam strong enough to ionize air at the focus is also strong enough to hurt an eye anywhere along its path, and it tends to glow as a streak rather than a point. So the design I computed uses two beams instead, crossing at an angle, each individually too weak to spark anything. Along their separate paths nothing happens, and only where they intersect does the combined field cross the ionization threshold.</p><p>This works because air-breakdown is one of the most nonlinear processes in ordinary physics. The ionization rate scales roughly as intensity to the eleventh power, so doubling the intensity raises the rate by a factor of two thousand. In my computed geometry, the crossing point ionizes about 10<strong>²³ times faster</strong> than anywhere along either beam. That is not a gentle preference but a switch: the glowing dot is pinned to one point in space because the physics essentially forbids it from appearing anywhere else.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VRsM_zZqpdQmd5RweUfEKw.png" /><figcaption><em>The localization, computed. As the two beams cross at a wider angle, the ignition core collapses from a smeared streak into a tight round point a few micrometres across. This is the difference between a glowing line and an addressable pixel.</em></figcaption></figure><p>A second choice makes it cheap. Instead of brute-forcing the air with raw power, you tune the two colours to land on real quantum energy levels in oxygen and nitrogen, pushing a swing at its natural rhythm rather than shoving it. This resonant scheme, a (2+1) REMPI ladder, drops the required intensity by about a factor of eighteen. The wavelength recipe fell out of the quantum chemistry: 2 × 239 nm + 727 nm for oxygen, 2 × 195 nm + 437 nm for nitrogen.</p><p>One detail there is worth dwelling on, because no amount of intuition would have produced it. The nitrogen channel that looks geometrically obvious, the one you would pick by inspection, turns out to be two-photon-dark. Forbidden, contributing nothing. You only discover that by computing the two-photon transition strengths; the obvious answer was wrong, and the only way to know was to do the arithmetic.</p><p>Assembled, the whole device is this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f4GAMfZ_uJ9ntZ3maqtw6Q.png" /><figcaption><em>Two sub-breakdown beams of different colours, one crossing point, and a scanner. The delivery beams stay individually below the ionization threshold; only the intersection lights up.</em></figcaption></figure><h3>The laser is three to eight times too expensive</h3><p>I priced a prototype from real vendor catalogues. It comes to roughly <strong>$378,000</strong>, and more than half of that is a single component: the femtosecond laser.</p><p>So I asked what the physics actually demands, rather than what a comparable system happens to use. A visible voxel needs about 50 nanojoules per flash, at roughly 100 femtoseconds, at about a megahertz repetition rate — around 50 milliwatts of average power. A laser pointer is five.</p><p>The industrial chirped-pulse amplifier at the heart of the $378k budget is therefore over-specified by three to eight times. What the design needs is a compact fibre MOPA, the kind of source that costs somewhere between $15,000 and $60,000.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/612/1*Ai8YC2uRfBEJHzwRAF0ZTg.png" /><figcaption><em>The single most expensive assumption in the build, priced three ways.</em></figcaption></figure><p>That one line item is most of the difference between a quarter-million-dollar research curiosity and something a well-funded lab could put on a bench.</p><p>I want to be careful about what I’m claiming here. I’m not saying anyone was foolish. When you assemble a system from the top down — pick a laser known to produce air breakdown, then build around it — you inherit that laser’s specifications as an assumption and never revisit them. Computing the device from the bottom up, starting from how many photons a dark-adapted human eye actually needs, produces a different answer.</p><p>That number is about 100 photons at the cornea, the Hecht–Shlaer threshold measured in 1942. My fiducial voxel delivers around 184, and the margin is conservative, because I modelled only the optically-thin continuum and ignored line emission that would make it brighter.</p><p>A related result was less expected. Voxel brightness climbs steeply with plasma <em>temperature</em>, from 107 to 282 photons per flash as the electron temperature goes from 1 to 3 eV, at almost fixed energy cost. Hotter plasma, not more energy, is the cheap brightness knob — a design lever that stays invisible unless you compute the emission spectrum.</p><h3>How it was built</h3><p>None of the above came from a single program. It came from about forty separate simulations, each in its own Docker container with its own pinned toolchain, orchestrated by an AI agent loop. Concretely, that loop ran in <strong>Claude Code</strong>, Anthropic’s terminal agent, on the <strong>Opus 4.8</strong> model. It planned the experiments, wrote the simulator input decks, ran the containers, read the resulting figures back, and argued with me about whether the numbers were physical.</p><p>The git history says the bulk of that work, 66 of 88 commits, landed in three days. Stretch it to six working days and you have essentially the whole study; the remaining calendar time was writing, a licensing pass, and gaps where I wasn’t touching it at all. What makes that possible isn’t speed of typing, but the collapsing cost of asking a well-posed physics question and getting back a validated numerical answer. The agent writes the input deck, runs the container, reads the plot back, and flags it when the answer disagrees with the textbook. My job became choosing which question was worth asking next, and refusing to accept any answer that hadn’t been checked against something real.</p><p>The architecture rule was Docker-per-simulator: no scientific dependency installed on the host, ever. Quantum chemistry in one container, particle-in-cell plasma codes in another, line-by-line radiative transfer in a third, custom Lagrangian radiation-hydrodynamics solvers in a fourth. A dozen tools spanning regimes that barely share vocabulary, let alone data formats.</p><p>The hard part of a project like this is not running any one simulator. It’s the bookkeeping across regime boundaries: units, validity ranges, which approximation is currently in force, what it means to hand a plasma state from a kinetic code to a chemistry solver.</p><p>So the station enforces a data contract. Every physical quantity is unit-typed with pint, so an energy can&#39;t be silently misread as a different unit three containers downstream. Results serialize to xarray datasets written as netCDF, and that file becomes the explicit input to the next stage, which makes every cross-regime coupling auditable rather than assumed. Every run is tracked in MLflow, keyed by configuration hash and container pin, so any figure in the paper traces back to the exact parameters and code version that produced it.</p><p>The agent layer sits on top with explicit roles: a planner that proposes the next configuration, a coder that writes simulator-specific input decks, a runner that executes them, an analyst that loads results and makes plots, and a critic that checks units, convergence, and physical plausibility before deciding what to run next. Each simulator is wrapped as an MCP server, so the same tool surface works regardless of which language or runtime lives inside the container. The repository also carries a fully local variant of the loop — the same planner-tool-analyst cycle driven by a Qwen model on the workstation with no API key at all — which exists to prove the architecture doesn’t depend on any one vendor.</p><p>One design decision mattered more than I expected: the figures get read back. Generated plots are inspected as part of analysis rather than treated as finished artifacts, and several discrepancies between what I expected and what the physics actually produced were caught by looking at a light curve and thinking <em>that shape is wrong</em>.</p><h3>The run I had to retract</h3><p>At one point I had a radiation-hydrodynamics simulation that reproduced the observed rise-and-decay shape of a real air-spark light curve. Beautiful agreement, exactly the kind of figure you build a section around.</p><p>It was wrong. The solver wasn’t conserving energy, and when I rebuilt it properly a spurious amplification factor collapsed from 1407× to 1.4×, taking the agreement with it. The match had been an artifact of the bug. The real prompt glow turned out to be continuum emission, and the light-curve shape had to be reassigned to line and non-LTE physics entirely. I withdrew the original conclusion rather than defending it, and it sits in the paper described as what it was.</p><p>This matters more than the result it replaced. A simulation campaign that never retracts anything isn’t careful, it’s unfalsifiable.</p><blockquote>The reason I trust the rest of the numbers is that this one failed loudly enough to catch, and the infrastructure was honest enough to let me catch it.</blockquote><p>That’s also why every result carries its caveats inline. TDDFT vertical energies come with ±0.2–0.4 eV uncertainty for nitrogen and ±0.5 eV for the multireference oxygen states, so the robust conclusion is the level <em>ordering</em> and the threshold <em>reduction</em>, not any absolute line position. Where a closed form or a real measurement exists, the computed number gets checked against it: nitrogen polarizability lands at 11.54 a.u. against a reference 11.74, a 1.7% error; a microwave breakdown field of 29.7 kV/cm recovers the textbook 30; a maser inversion threshold matches its analytic form to machine precision.</p><p>I also ran a negative control on purpose. There’s a tempting shortcut: skip the plasma, use a lot of cheap laser pointers, and let ambient dust scatter the light into a visible point. I computed it, and it cannot work. Scattering is linear, so power scales with intensity and a beam glows along its entire path rather than at a point; crossing N beams raises the junction only N-fold instead of by 10²³. Ordinary air has far too little dust anyway, and the side-on luminance comes out dark-room-marginal at best.</p><p>That failure is informative, because it is precisely why the nonlinear plasma approach is the right one. I’d rather publish the dead end than quietly not mention it.</p><h3>What this is, and what it isn’t</h3><p>Two things need saying plainly, because the word “hologram” does a lot of unearned work here.</p><p>It isn’t a hologram, technically. A real hologram reconstructs an image from interfering light waves, whereas this makes light <em>at</em> each point by igniting the air there. The everyday word fits what you would see, but it’s the wrong word for how it works.</p><p>And it isn’t built. Every number in the study is computed, and nobody has put this design on a bench and measured it. The models are reduced in places, the budgets are representative, and no physical device has ever been compared against them. So the conclusion is not that it works. The conclusion is:</p><blockquote>There is no remaining physics reason it shouldn’t. What’s left is engineering and measurement, not open questions.</blockquote><p>Every link in the chain holds. The spark is bright enough to see; the beams can be held on target to about 1.3 micrometres by a fast-steering mirror and a camera watching where the dots actually land; the image is safe to touch; the ozone and nitrogen oxides it produces stay well inside health limits with modest ventilation; and the whole thing runs on sub-watt average power. The beams themselves are still Class-4 and need to be boxed inside the image volume and certified, which is a known engineering job rather than a mystery.</p><p>What I genuinely hope someone does with this is build a small one and tell me where I’m wrong. And if it does work, I wouldn’t say no to a line in the acknowledgements. <em>“Contributed the design study; never turned on a laser”</em> would be entirely accurate.</p><p>Until then I’ll settle for the one thing I have unambiguously earned, which is the ability to say <em>I designed a hologram</em> and let people fill in the rest themselves. Andrej Karpathy gave the world <em>vibe coding</em>. If any of this survives contact with an actual optical bench, I’d like to lodge an early claim on <em>vibe science</em> — with the caveat, which is the entire point of this article, that every number in it was cross-checked against published literature or an analytic limit, and the one result that couldn’t survive that got retracted.</p><p><em>The full manuscript, the LaTeX source, the simulation station, and every container definition are open — code under MIT, paper and figures under CC BY 4.0: </em><a href="https://github.com/LumiVerseHR/air-laser-physics"><em>github.com/LumiVerseHR/air-laser-physics</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a365df88b538" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Great AI Inference Speed Race: Is it possible to run LLMs at 2000T/s]]></title>
            <link>https://medium.com/@tihomirjauk/the-great-ai-inference-speed-race-is-it-possible-to-run-llms-at-2000t-s-3ea4cc4f1a0d?source=rss-d06d4e5831c0------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ea4cc4f1a0d</guid>
            <category><![CDATA[cerebras-ipo]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm-inference]]></category>
            <category><![CDATA[togetherai]]></category>
            <dc:creator><![CDATA[Tihomir Jauk]]></dc:creator>
            <pubDate>Sun, 20 Jul 2025 17:45:04 GMT</pubDate>
            <atom:updated>2025-07-20T17:45:04.647Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ysdxH8eUc8NLOqQA" /><figcaption>Photo by <a href="https://unsplash.com/@vantaymedia?utm_source=medium&amp;utm_medium=referral">Van Tay Media</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>As an AI developer constantly working with large language models, I’ve been curious about what would another order of magnitude in inference speed enable us to do. Recent buzz around specialized inference chips like Cerebras’ Wafer-Scale Engine had me wondering: just how much faster could AI inference get?</p><p>The question that sparked this investigation was simple: <strong>Which provider actually delivers the fastest inference for Llama-3.3–70B, and at what cost?</strong></p><p>To find out, I decided to put four major providers head-to-head: Together.ai (my longtime favorite), SambaNova, GroqCloud, and Cerebras.</p><p>I used the same prompt across all providers to generate a comprehensive article about AI inference speed (meta, right?). The prompt was designed to produce substantial output — around 1,500 words — to test sustained generation performance rather than just quick responses.</p><p>Here’s what I measured:</p><ul><li><strong>Tokens per second</strong>: Raw generation speed</li><li><strong>Total latency</strong>: End-to-end completion time</li><li><strong>Pricing</strong>: Cost per million tokens</li><li><strong>Real-world performance</strong>: How it feels to use</li></ul><h3>The Results: Speed Showdown</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XlvrnYXyiUGxasRAAl4HuA.png" /></figure><p>Cerebras didn’t just win — it <strong>obliterated</strong> the competition with 1,643 tokens per second, nearly <strong>4.3x faster</strong> than the nearest competitor and <strong>9.6x faster</strong> than Together.ai.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YvFEJJQcLGazXLlG0BHu5Q.png" /></figure><p>But speed is only part of the story. Let’s look at the economics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ezbMK1tMTd0YXSDpNICwdw.png" /></figure><p><em>*Prices per 1 million tokens<br>**Blended price assumes 3:1 ratio (3 input tokens for every 1 output token)</em></p><p>Here’s where it gets interesting: Cerebras not only delivers the fastest performance but also offers the <strong>most cost-effective pricing</strong> at $0.60 per million tokens blended. This creates an incredible value proposition.</p><h3>Performance vs Price</h3><p>When we calculate the speed-to-cost ratio (tokens per second per dollar spent), the results are clear:</p><ol><li><strong>Cerebras</strong>: 2,738 T/s per $1M (Best value)</li><li><strong>GroqCloud</strong>: 535 T/s per $1M</li><li><strong>SambaNova</strong>: 295 T/s per $1M</li><li><strong>Together.ai</strong>: 225 T/s per $1M</li></ol><p>Cerebras delivers <strong>12x better value</strong> than Together.ai when considering both speed and cost.</p><h3>The Fun Math: Daily Spending at Maximum Speed</h3><p>Just for entertainment, I calculated what would happen if you ran each provider at maximum speed for 24 hours straight:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5t5sP2zh_ebUdeyUBcRI0A.png" /></figure><p>Cerebras would generate enough tokens in one day to write approximately <strong>2,132 novels</strong>! While the daily cost is higher, you’re getting an astronomical amount of processing power.</p><h3>My Experience: From Together.ai Fan to Cerebras Convert</h3><p>I’ll be honest — I went into this test as a Together.ai advocate. Their platform has been my go-to for months due to its excellent developer experience, comprehensive model selection, and reliable performance. Together.ai excels at providing a stable, well-rounded service that just works.</p><p>But Cerebras completely changed my perspective on what’s possible with AI inference. Responses that would take 7+ seconds on other platforms appear almost instantly on Cerebras. It’s the difference between waiting for AI and having AI keep up with your thinking.</p><p>The secret lies in their Wafer-Scale Engine (WSE-3), a revolutionary chip design that processes AI models on a single, massive wafer rather than connecting multiple smaller chips. This architecture eliminates the bottlenecks that plague traditional GPU-based inference.</p><h3>Real-World Implications</h3><p>This isn’t just about bragging rights. Ultra-fast inference enables entirely new categories of applications:</p><p><strong>Conversational AI</strong>: Real-time, natural conversations without awkward pauses</p><p><strong>Code Generation</strong>: Instant feedback and suggestions in development environments</p><p><strong>Content Creation</strong>: Rapid iteration on writing, marketing copy, and creative content</p><p><strong>Agentic Workflows</strong>: AI agents that can think and respond as fast as humans expect</p><h3>Claude Code + Cerebras = A Developer’s Dream</h3><p>As someone who uses Anthropic’s Claude Code extensively for development work, I can’t help but imagine the possibilities of combining Claude’s reasoning capabilities with Cerebras’ speed. Currently, Claude Code, while incredibly powerful, can sometimes feel slow when processing complex codebases or generating substantial amounts of code for analysis and planning.</p><p>Anthropic should seriously consider partnering with Cerebras to accelerate Claude Code. Imagine having Claude’s sophisticated reasoning and code understanding delivered at 1,600+ tokens per second. It would transform the development experience from “waiting for AI assistance” to “AI keeping pace with developer thinking.”</p><p>Such a partnership could make Claude Code the fastest and most capable coding assistant available, setting a new standard for AI-powered development tools.</p><h3>The Verdict</h3><p>Cerebras has fundamentally shifted my understanding of what AI inference can be. The combination of blazing speed and competitive pricing creates a compelling value proposition that’s hard to ignore.</p><p>For developers and companies serious about AI performance:</p><ul><li><strong>Choose Cerebras</strong> for maximum speed and best value per token</li><li><strong>Choose GroqCloud</strong> for balanced performance and pricing</li><li><strong>Choose Together.ai</strong> for ecosystem richness and reliability</li><li><strong>Consider SambaNova</strong> for specific enterprise requirements</li></ul><p>The future of AI isn’t just about smarter models — it’s about making those models instantly responsive. Cerebras is leading that charge, and I suspect we’re only seeing the beginning of what ultra-fast inference will enable.</p><p><strong>What’s your experience with different inference providers? Have you tried Cerebras yet?</strong> I’d love to hear about your own speed tests and use cases in the comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3ea4cc4f1a0d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>