Playing around with GPT-2 - the sequel

the sequel

04-01-2022

Contents

When I <a href="/story/playing-around-with-gtp-2/">played with the smaller GPT-2 models</a>, the output was fun but obviously a machine talking to itself — grammatical, topically vague, and prone to collapsing into nonsense after a sentence or two. OpenAI had deliberately held back the full model, worried it was too good and too easy to abuse for spam and fake news. On 5 November 2019 they changed their minds and released the lot: the full 1.5-billion-parameter model, <code>1558M</code>, having “seen no strong evidence of misuse so far”. So of course I ran it against exactly the same prompts, to see how much the extra parameters actually buy you. The short answer: a lot. The difference between the small model and the full one isn’t subtle. Where the little models produced word-salad that happened to be grammatical, <code>1558M</code> holds a thread across a whole paragraph, keeps its pronouns straight, and every so often says something that would pass a lazy skim. It’s the first time one of these felt less like a Markov chain with delusions of grandeur and more like something that had genuinely read the internet. A note from 2026: this post is a period piece. The commands below use TensorFlow 1.x and the original <code>openai/gpt-2</code> repository, which is how you ran these things in 2019. If you want to run a small language model on your own hardware today the toolchain is completely different and far easier — see <a href="/story/running-gemma-3-locally-googles-small-model-on-consumer-hardware/">running Gemma 3 locally</a> for the modern equivalent. I’ve left the original workflow intact because the outputs are the interesting part, and they’re a nice fossil of where this technology was. <h2 id="downloading-the-full-model">Downloading the full model</h2><div class="ad-unit ad-in-article" aria-label="Advertisement"> Advertisement <ins class="adsbygoogle" style="display:block;text-align:center" data-ad-client="ca-pub-3726833845844946" data-ad-slot="3291553914" data-ad-format="auto" data-full-width-responsive="true"></ins> <script>(adsbygoogle = window.adsbygoogle || []).push({});</script> </div> If you followed the guide in the <a href="/story/playing-around-with-gtp-2/">original article</a>, the new weights drop straight into the same checkout: <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code>1 </code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-shell" data-lang="shell">python3 download_model.py 1558M</code></pre></td></tr></table> </div> </div> The model is roughly 6 GB, and at the time the download seemed throttled, so it takes a while. Once it’s down, point the interactive sampler at it: <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code>1 </code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-shell" data-lang="shell">python3 src/interactive_conditional_samples.py --model_name=1558M --top_k 40 --nsamples=2 --length 100</code></pre></td></tr></table> </div> </div> <code>--top_k 40</code> restricts each step to the 40 most likely next tokens, which keeps the output coherent rather than letting a rare low-probability token derail a sentence; <code>--nsamples=2</code> gives two completions per prompt so you can see the variance; <code>--length 100</code> caps each at 100 tokens. Now the same prompts as before. <h2 id="the-papers-own-title">The paper’s own title</h2> Feeding it the title of the paper that introduced GPT-2, the full model produces something that reads like a plausible (if invented) machine-learning abstract rather than the small model’s vague waffle: <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 </code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-shell" data-lang="shell">Model prompt >>> Language Models are Unsupervised Multitask Learners. ======================================== SAMPLE 1 ======================================== We propose a neural language model (NLLM) for multistep learning that allows models to learn more than one state simultaneously. We focus on the sequential aspect of recurrent neural networks and show that this can provide a powerful approach to learning multistep tasks, such as speech recognition. In our model, neural processes are represented in the form of a vector, which we feed into the network. Each network layer is a function that takes a vector and returns a matrix containing the ======================================== SAMPLE 2 ======================================== This post is part of a series on Neural Networks with Deep Belief Networks. We describe our architecture and discuss how we use it in our latest project: C2. When your app starts up, it loads many different modules from different resources. You may need to keep an eye on the models and their dependencies in the app. These models depend on one another, so you may need to watch many of them at once. A typical scenario would look like this: Model A ================================================================================</code></pre></td></tr></table> </div> </div> Notice it invents a plausible-sounding acronym (“NLLM”), frames a research direction, and even gestures at citations. None of it is real, but the shape of a technical paper is there, which is exactly the failure mode people were nervous about: fluent, confident, and wrong. <h2 id="the-star-wars-question">The Star Wars question</h2><div class="ad-unit ad-in-article" aria-label="Advertisement"> Advertisement <ins class="adsbygoogle" style="display:block;text-align:center" data-ad-client="ca-pub-3726833845844946" data-ad-slot="3291553914" data-ad-format="auto" data-full-width-responsive="true"></ins> <script>(adsbygoogle = window.adsbygoogle || []).push({});</script> </div> Ask it the eternal fandom flamewar — did Han shoot first? — and the full model is markedly more fluent than the small one, though it wanders off into what reads like a translated web-novel: <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 </code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-shell" data-lang="shell">================================================================================ Model prompt >>> Did Han shoot first ? ======================================== SAMPLE 1 ======================================== Han shot first, so he should be the one who would be punished, right ? But… No way… the judge should be the other party! "You are right, Han shot first." An officer with a serious expression told Feng Nu beside her, his face expressionless as before. Feng Nu shook her head: "It really didn't matter, Judge, Han is a military officer now. The rules are the same. I do understand ======================================== SAMPLE 2 ======================================== That is a very interesting question, and one I am not sure about myself. To my knowledge, Han shot his first-round shot first. However, at the range, I am certain Han hit another target on at least two different occasions. I also think his point shot (from his second shot) was the only shot that hit on target. Why doesn't he have more of these marksmanship feats? I can only speculate, but here are the possible ================================================================================</code></pre></td></tr></table> </div> </div> The “Feng Nu” and “Kirito” names are borrowed wholesale from anime and light-novel training data — a good illustration that the model has no idea what it’s talking about, it’s just steering towards whatever text-world its opening tokens landed in. It committed to a courtroom drama and furnished it from the nearest available genre. <h2 id="the-avengers-plot">The Avengers plot</h2> Asked for the plot of “the new Avengers movie”, it produces confident-sounding production trivia — a named director, co-writers, a city — none of it a real plot, but every detail plausibly formatted: <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code> 1 2 3 4 5 6 7 8 9 10 11 </code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-shell" data-lang="shell">================================================================================ Model prompt >>> What is the plot of the new avengers movie? ======================================== SAMPLE 1 ======================================== pic.twitter.com/v5pVZW1q5B — Kevin McCorry (@K_McCorry1) March 18, 2016 The new avengers is being directed by Joss Whedon (Dr. Horrible's Sing-Along Blog, Buffy the Vampire Slayer) who co-wrote the upcoming Avengers sequel film with Mark Millar (Kick-Ass, Preacher). It follows the new crop of "Avengers" heroes and villains as they form an unlikely ======================================== SAMPLE 2 ======================================== We know the new one takes place in Chicago, but what was the city at the first movie? CRAIG ANTHONY: We had a very open conversation with all our vendors, the way Marvel has their licensing. We asked them to let us know what they would want in a city, and what would their top five choices be, and how they would build a city. We did some extensive research. We went on a map; we went through every street on our own, ================================================================================</code></pre></td></tr></table> </div> </div> It confidently attributes the film to Joss Whedon and Mark Millar, complete with parenthetical CVs. This is the model’s most consistent trick: it has absorbed the format of entertainment journalism so thoroughly that it can generate press-release prose indefinitely, whether or not any of it corresponds to reality. <h2 id="brexit">Brexit</h2> On Brexit the full model is, appropriately, unsure of the exact mechanics — but it does something the small model never managed, which is to construct a genuinely two-sided argument across the two samples: <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 </code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-shell" data-lang="shell">================================================================================ Model prompt >>> How about brexit? ======================================== SAMPLE 1 ======================================== We have a Brexit option, that's fine. A no-name "Brexit" option - a bit like the way the euro has been handled by some countries, with some people saying "it wouldn't be as bad as the euro, wouldn't be a catastrophe, would be fine. And then they go on to do it. The problem with that though, is you want people to vote against their own interests, and if you start to take away options that you want people ======================================== SAMPLE 2 ======================================== The UK's exit from the EU would not immediately lead to a hard border with Irish Republic, but the country's border with Northern Ireland is already a significant bottleneck crossing into either side of the border at Belfast. A hard border would have to be agreed after Brexit. Image copyright AFP What happens in a no-deal Brexit? It is not entirely clear what would happen in a no-deal Brexit. The EU would need to agree a transition deal to avoid a ================================================================================</code></pre></td></tr></table> </div> </div> The second sample even reproduces the “Image copyright AFP” furniture of a news page and pivots into the Northern Ireland border question — the actual crux of the whole negotiation. It doesn’t understand it, but it has clearly ingested enough political reporting to land near the real issues, which the small model never did. <h2 id="the-2020-election">The 2020 election</h2> Asked whether Trump would be re-elected in 2020, the model is fairly sure he would, frets about the treatment of women in the administration, and — for reasons known only to its training corpus — free-associates its way into a Game of Thrones reference: <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 </code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-shell" data-lang="shell">================================================================================ Model prompt >>> Will donald trump be reelected in 2020 ? ======================================== SAMPLE 1 ======================================== What role will women play in his administration ? Who will be running his administration when they leave it ? We don't know the answers to many of these questions. And if we go by the conventional wisdom, the answer to none. A friend of mine, an independent thinker with some interest in politics, just asked me, "Who are the three people he hates the most?" The answer: None. The list includes the President and Vice president. That's a good question. Why ======================================== SAMPLE 2 ======================================== I think so. Who's your "dream candidate" ? A political philosopher or a social worker ? Politicians are like sex. It's really just one way of looking at it. Follow @SethJames on Twitter.<|endoftext|>Ahead of the release of the long-awaited season 5 of Game of Thrones, it seems we may have a few secrets to add to our list of suspects before we begin guessing who will be the Night King's victims. According ================================================================================</code></pre></td></tr></table> </div> </div> The <code><|endoftext|></code> token in the middle of that sample is the model hitting its document boundary and simply starting a fresh, unrelated document — in this case a Game of Thrones recap. It’s a useful reminder that GPT-2 isn’t answering your question; it’s predicting the most likely continuation of a text stream, and sometimes the most likely continuation is “this article ended and a completely different one began”. <h2 id="what-this-told-me">What this told me</h2> Two things stuck with me after an afternoon of this. First, the jump from the small models to <code>1558M</code> is not linear. The extra parameters don’t just make the output a bit tidier — they cross a threshold where the text stops being obviously generated on a casual read. That’s the whole reason OpenAI hesitated, and running both models back to back you can feel exactly what worried them. Second, and more important: coherence is not comprehension. Every one of these samples is confidently, fluently wrong in ways that a distracted reader would never catch. It invents directors, fabricates citations, and states falsehoods in the exact register of authority. The lesson that has aged best from 2019 is that the dangerous property of these models was never that they’d be unconvincing — it’s that they’d be too convincing while having no relationship to the truth. Everything that came after, from GPT-3 to today’s assistants, is an escalation of precisely that tension, and it’s why “sounds right” and “is right” remain two entirely different things worth keeping separate in your head. If you want to see where the small model landed on other topics, the <a href="/story/playing-around-with-gtp-2/">original write-up</a> has it wandering through bitcoin, government shutdowns, the Mediterranean diet, and the mighty ibex.

Written by Smarc

Founder and editor of vo.rs. A lifelong tinkerer who self-hosts far more than is sensible, hardens Linux boxes for fun, and prods the latest AI tools to see what they can really do. The how-to guides here are the notes Smarc wishes had existed the first time round.

Tagged#ai #machine-learning #gpt-2 #natural-language-generation

Playing around with GPT-2 - the sequel

the sequel

Related Content

Fine-Tuning vs Prompting vs RAG: Picking the Right Tool Without Wasting GPU Hours

Talking to Your Documents: A Practical RAG Pipeline with Open-Source Tools

Prompt Injection: The SQL Injection of the AI Era

What Is Agentic AI, and Why Is Everyone Suddenly Talking About It?