Talk by Laurie Voss (Arize AI) on instruction-following capacity limits in LLM agents, measured with the IFScale benchmark, which asks a model to write a business report containing a list of exact words and counts how many appear. A year ago frontier models began dropping instructions around 200–300 items, capping skills-file size; rerunning the test today, current models scored 100% at the old scale, and the benchmark had to be raised from 500 to 10,000 words before failures appeared. The limit now sits near 2,000 instructions (about 5,000 for the best model), roughly a tenfold gain in a year. Notably, models now fail in different ways: one simply forgets rules, one refuses at the API level when random words trip its safety classifier, one exhausts its reasoning budget verifying instructions, and one abandons the task mid-report. Voss argues the core problem has shifted from fitting rules into a small budget (compression) to verifying whether the model actually obeyed them (output checking). Covers the benchmark's design, replicating last year's results, the tenfold jump, failure modes, workflow implications, and caveats about what the test does not prove about reliability. # Channel Presentation Observations — Laurie Voss, Arize AI ## Speaker & Channel Identity - **Explicit self-description**: The speaker identifies himself as **Laurie (Lori) Voss, "head of developer relations at Arize AI"**, and notes he "co-founded npm Inc." in a former career. > "I'm Lori. I'm head of developer relations at Arise AI. Uh in a former life, I co-founded npm Inc." - **Observable presentation**: This is a **conference talk format** ("Thank you for coming to this delightfully nerdy talk"), delivered to a live audience, framed with time-boxed promises ("my promise for your next 18 minutes"). ## Intended Audience - **AI application developers / engineers** who write "skills files" and agent prompts — evidenced by repeated second-person workflow framing: > "how long your skills and prompts can actually be and what changes you should make to your workflow as a result." > "Why should you care when you get back to your desk?" - Audience is assumed technically fluent: terms like **API-level responses, evals, token budgets, log scale, benchmarks** are used without explanation. - Audience is expected to be current on AI models (GPT 5.5, Claude Opus 4.7, Gemini 3.1 Pro, DeepSeek V4 Pro are named without introduction). ## Claimed Scope of Content - **Empirical benchmark research**: The talk claims to replicate and extend the **IFScale benchmark** (instruction-following capacity), reporting a roughly **10x improvement** in frontier-model instruction capacity in ~12 months. > "in about 12 months frontier models got close to 10 times better at following instructions simultaneously." - **Failure-mode taxonomy**: Claims four distinct failure modes per model family (Deepseek forgets; Claude refuses on safety filters; Gemini exhausts thinking tokens; GPT 5.5 quits mid-report). - **Practical guidance**: Claims three workflow changes follow from the findings, culminating in a "verification problem" thesis: > "the compression problem is gone... The new hard part is knowing whether it actually did what you said and that is a verification problem." ## Demonstrated Presentation Patterns - **Transparency about method and limitations**: - Cost disclosure: "It cost me $29 to run all of these queries. 2,300 calls across seven models." - Caveats given proactively: "this is a proxy task... It is not the same as proof that a long skills file works." - Data/code sharing: "If you want all of the code and all of the data, uh it is at this GitHub URL." - **Self-aware humor and informality**: "which is kind of on brand, isn't it? \[snorts\] uh which, you know, I would never say that out loud"; the talk is billed as "delightfully nerdy." - **Acknowledges employer tie-in with restraint**: "there are plenty of plugs for Arize. So I'm just going to say one true thing..." — the product mention is folded into the verification argument. - **Cites external research**: References Chroma's "context rot" work, and a 46-model paper "revisiting the reliability of language models in instruction following," plus new benchmarks (Firebench, CCR bench, Guidebench), positioning the talk within an emerging research conversation. ## Distinguishable Claims vs. Demonstrations - **Creator claim (stated)**: Novelty and authority — "I did some research and then a whole bunch of other actual scientists piled in"; "this is the last possible time that I could have run this test." - **Demonstrated (observable)**: A structured empirical method (replication of prior benchmark, log-scale results, model-by-model failure analysis), a pedagogical delivery style with concrete numbers, and a developer-relations posture that blends research findings with actionable advice. - **No private motive inferred**: The speaker's affiliation with Arize AI is explicitly disclosed; product advocacy is limited to the quoted evals passage and a marketing-inserted QR code for a social event, which the speaker flags as such. # Delivery Observation — Laurie Voss, "How long can your skills be before your agent forgets what you told it?" ## Tone & Overall Style - **Conversational, self-deprecating, humor-forward delivery.** Opens with audience rapport: *"Hello everybody. Thank you for coming to this delightfully nerdy talk."* - Uses **wry asides** throughout: *"which is kind of on brand, isn't it? Um [snorts] uh which, you know, I would never say out loud."* - Ends on a deliberately anticlimactic comedic beat: *"they might already be wrong. Boom. Be wrong."* ## Filler & Disfluency Pattern - **Heavy, consistent use of "uh" and "um"** across the entire talk — this is a stable, pervasive feature, appearing in nearly every transcript segment (e.g., *"Uh this talk has a really long title. Uh so let me give you the short version up front."*). - Combined with **false starts and self-corrections**: *"What our test doesn't do is measure whether…"* and *"and Gemini DPT. this is a a stupid thing to ask for"* — suggesting a spoken, unscripted-feeling register. ## Vocabulary & Recurring Phrases - **"Skills file" / "instructions"** — the central recurring vocabulary, used dozens of times. - **"Ceiling"** as a dominant metaphor: *"that's a really hard ceiling"*, *"this benchmark was written to top out at 500 words"*, *"we'd built a test to find the ceiling and the models had walked straight through the ceiling"*, *"The ceiling moved by 10x in one year."* - **Failure-mode personification** — describes models with personality-like verbs and framing while keeping it observational: - *"Deepseek 4 is a traditional model. It just forgets things. It doesn't have any drama."* - *"Claude gets scared and refuses, Gemini overthinks itself into silence, and GPT 5.5 finishes half of the job and tells you that the rest of it is beneath it."* - **"On brand"** appears twice ("extremely onbrand way", "kind of on brand, isn't it?") — a recurring casual idiom. ## Teaching Habits & Structure - **Explicit roadmap/promise-setting**: *"So here's my promise for your next 18 minutes. Uh, I'm going to show you where that 200 number came from, whether it's still true, and what the real number is today."* - **"Money slide" framing**: *"this is the money slide. This is the results."* - **Numbered takeaways**: *"three things have changed to your workflow. The first is… number two is… but the third thing is the big one."* - **Explicit caveats section preempting Q&A**: *"and now some caveats uh to head off the Q&A. Um, first and important first and most important…"* - **Concrete, vivid hypotheticals** to make abstract instructions tangible: *"if the user says X, do Y, always include a section on Z, never use the phrase W."* ## Rhetorical & Audience-Engagement Patterns - **Rhetorical questions**: *"did it actually pay attention? Did it actually follow all of these rules or did it just sort of, you know, do what it felt like?"*; *"You can't really tell or can you? More on that later."* - **Direct audience appeal to shared experience**: *"you know the feeling that I'm talking about"*; *"I don't know if everybody else feels this way but…"* - **Anticipated objection, then concession**: *"now you might say uh that including random words in a report is not the same as following real instructions and fair enough"* — a repeated move ("it is not the same as proof that a long skills file works"). ## Narrative & Self-Positioning Choices - **Story-driven framing of research origin**: builds the talk around a conference anecdote (*"I was watching a talk by Dexter Horthy… he mentioned as an aside…"*). - **Recurring "science" framing**: *"before chasing new models you have to do good science which means that you have to replicate the uh old result"*; later contrasts himself with *"actual scientists"*: *"I'm not a scientist. I did some research and then a whole bunch of other actual scientists piled in."* - **Light vendor disclosure with self-awareness**: *"you knew I was going to mention evals eventually because I work at Arise and this is where I do that. Um, but there are plenty of plugs for Arise. So I'm just going to say one true thing."* ## Uncertainty Noted - The transcript reflects spoken delivery only; no visual channel is available, so observations about slides, pacing, or gesture are limited to what is verbally referenced (e.g., "money slide," log-axis explanations, QR codes). - Filler-word frequency and self-corrections may be inflated or distorted by transcription artifacts; the *presence* of these features is reliable, exact counts are not. # Observation Document: Skill File Instruction Limits **Video:** "How long can your skills be before your agent forgets what you told it?" — Laurie Voss, Arize AI ## Speaker - **Laurie Voss**, Head of Developer Relations at Arize AI; co-founder (former) of npm Inc. Sole speaker; positions below are all his. ## Origin of the Question - At AI Engineer in Miami, **Dexter Horthy** mentioned (as an aside) that an agent can follow about **200 instructions** before forgetting them, citing a figure from 2025 — per Voss, Horthy noted things might be better now. - Voss's reaction: 200 is very few — "a decent skills file blows past 200 instructions almost immediately" — and would be a hard ceiling on buildable complexity. ## The Benchmark (IFScale) - Source: a paper by "Jeroslowitch" (Voss flags he may be mispronouncing) and co-authors. - Method: ask a model to write a business report including N exact required words; each word = one instruction. Measures **density (n)** vs. **accuracy**. - Voss's qualification: it's a **proxy task**, not real instructions — but "include the word revenue is the same shape of task as include a section on pricing." If a model can't track 200 words, it will do **worse** with 200 complicated instructions, so 200 is a **ceiling**. ## Replication (Baseline) - Only 3 of the original 10 models remained API-available: **GPT-4.1, Claude Sonnet 4, Gemini 2.5 Pro** (one has since been retired). - Results matched the original paper within noise: **a year ago, ~200–300 rules** is where frontier models started failing; by 500 rules, 30–50% lost. ## Current Frontier Results - Tested: **GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro, DeepSeek V4 Pro**. All aced the original test, forcing Voss to extend the vocabulary from 500 → 10,000 words. - **Headline: instruction capacity improved ~10x in 12 months** — cliff now at ~2,000 instructions generally, up to **5,000** for the best models. - Voss notes Opus 4.8 (released after his test) is even better; "this chart is a little out of date already." ## Distinct Failure Modes - **DeepSeek V4**: quietly forgets — starts dropping around 750 rules, nearly half by 2,000. "The failure mode that I trust most." - **Claude Opus 4.7**: safety classifier triggers on word combinations (e.g., "anthrax and cyanide") and refuses at the API level. Voss had to pre-filter words through an OpenAI safety filter; after that Claude did well. Refuses early on medical-looking content. - **Gemini 3.1 Pro**: solid to 5,000 instructions, then exhausts its token budget on thinking and produces little/no usable output. - **GPT-5.5**: best overall — 99% accuracy to 5,000 rules — but when pushed, starts writing, then ~500 words in declares the test stupid and stops. Voss: it's "not wrong," but the half-finished output **looks like a real answer**, making it the hardest failure to detect. ## Workflow Implications (per Voss) 1. Skills files no longer need to be kept under 200 instructions with sub-skill labyrinths — **they can be long**. 2. Hundreds of specific rules can go directly in the prompt; 2,000 constraints ≈ an entire style guide that previously required sharding across agents. 3. The question shifted from "can the model do this?" to **"is it worth the cost/latency?"** — a hard wall became a soft trade-off. ## Caveats and Unresolved Tensions (stated by Voss) - Keyword inclusion is **evidence, not proof** that long skills files work. - Models hit walls at wildly different points (**750 to 9,000+**), so model choice matters. - His test does **not** measure whether models *reason clearly* over giant prompts. - **Chroma's "context rot" work** (18 models): accuracy on long inputs can fall 30–50% before the context limit; counterintuitively, well-structured text may fail *more* than shuffled instructions — Voss explicitly says "I don't know why that's the case." - **"Revisiting the Reliability of Language Models in Instruction Following"** (46 models): models can ace a benchmark yet be wildly unreliable — rewording or **reordering** the same instructions can radically change performance. Optimal ordering is still open research. **Capacity went up; reliability is still a problem.** - Claude's refusals are loud (safe); GPT's quiet half-finished output is dangerous — outputs must be verified every time. ## Conclusions - The **compression problem is gone**; the new problem is **verification**, which "doesn't get solved by writing a better prompt" — it's solved by evals/monitoring outputs with another LLM (which he notes is what Arize does). - Assumptions about prompt/skills-file size made **six months ago are probably already wrong** — re-engineer. - Side notes: total research cost **$29** across ~2,300 calls and seven models; newer benchmarks now exist (**Firebench, CCR bench, Guidebench**); code and data available on GitHub. # Speaker & Provenance Observation Record ## Video - **Title:** How long can your skills be before your agent forgets what you told it? — Laurie Voss, Arize AI - **Video ID:** `01a15ca7-8e1c-7066-825b-b1f4eea7d9c4` - **No curated metadata supplied**; provenance facts below come solely from the transcript. ## Speakers & Roles - **Laurie Voss (self-identified as "Lori")** — **Sole speaker / presenter**. - Self-described role: **Head of Developer Relations at Arize AI**. - Prior affiliation: **co-founder of npm Inc.** ("in a former life, I co-founded npm Inc."). - Explicitly disclaims scientific standing: > "I'm not a scientist. I did some research and then a whole bunch of other actual scientists piled in." - **No other speakers, guests, narrator, or panelists** appear in the transcript. ## Third-Party Claims & Referenced Sources - **Dexter Horthy** — cited talk at "AI Engineer in Miami"; aside claim that an agent follows ~200 instructions before forgetting (figure "from 2025"). - **IFScale benchmark** — paper by an author the speaker names as "Jeroslowitch" (speaker explicitly flags uncertainty about the name) and co-authors; source of the 200-instruction figure. - **Chroma's "context rot" work** — across 18 models; claims accuracy on long inputs falls 30–50% before the context window limit; finding that coherent structured text fails more than shuffled instructions (speaker notes he hasn't read the full report). - **"Revisiting the Reliability of Language Models in Instruction Following"** — paper testing 46 models; finding that rewording/reordering the same instructions changes compliance. - **Benchmarks:** Firebench, CCR bench, Guidebench — cited as measuring constraint-following. - **Models referenced as test subjects:** GPT 4.1, Claude Sonnet 4, Gemini 2.5 Pro (replication); GPT 5.5, Claude Opus 4.7, Gemini 3.1 Pro, DeepSeek V4 Pro (new tests). Speaker notes Opus 4.8 released after his testing and that one of the three legacy models was retired since publication. ## Speaker's Own Research Claims (provenance: presenter, unverified) - Replication matched original IFScale results within noise. - New frontier models handle ~2,000–5,000+ constraints; roughly 10x improvement in a year. - Four failure modes described: DeepSeek quietly forgets; Claude refuses at API level (mitigated by pre-filtering words through OpenAI's safety filter); Gemini exhausts thinking tokens; GPT 5.5 abandons tasks mid-output. - Cost claim: $29 for ~2,300 calls across seven models. ## Promotional Segments (Flagged) - **Arize AI / evals plug:** speaker states > "I work at Arise and this is where I do that... that is an eval. And that is what Arize does." (He acknowledges > "there are plenty of plugs for Arize.") - **Marketing QR code:** > "this other QR code is something marketing made me insert. We are having a World Cup watch party tonight at 5:00 p.m." — link to a Luma event, explicitly attributed to marketing. ## Availability & Access Notes - Code and data offered via a **GitHub URL** and **QR codes** (URLs not readable from transcript). - Only 3 of the original paper's 10 models remained API-available at test time; one of those retired since publication — availability limits on replication. - **Comment availability:** not stated in transcript or metadata; unknown. ## Uncertainty Notes - Original paper author's name flagged as likely mispronounced by speaker. - Speaker states his keyword task is a **proxy**, not proof, that long skills files work. All right. Hello everybody. Thank you for coming to this delightfully nerdy talk. Uh this talk has a really long title. Uh so let me give you the short version up front. You write skills files uh and stuff them full of instructions. At some point the model stops keeping track of all of them. The question is where is that point? At what point have you put too many instructions in your skills files? Uh and the answer has changed a lot in the last year. I'm Lori. I'm head of developer relations at Arise AI. Uh in a former life, I co-founded npm Inc. So some of you may know me from the days of JavaScript. These days I spend a lot of time thinking about AI and how to test it. Uh a few months ago I was at AI engineer in Miami which was a good conference. Uh and I was watching a talk by Dexter Horthy. Uh it was a good talk. It was not about this topic at all. Uh but while he was giving that talk he mentioned as an aside uh that an agent can follow up to about 200 instructions uh before it starts forgetting those instructions. Uh and he then he moved on in his talk and it was entirely an aside. Uh and he he mentioned that that figure is from 2025 so things might be better now. Um, and I stopped listening for a second because I was like, 200 instructions. Uh, is not very many instructions at all. Right? A decent skills file blows past 200 instructions almost immediately. Um, if the user says X, do Y, always include a section on Z, never use the phrase W, every one of those is a separate instruction. Uh, and if the model quietly stops tracking them after 200, that's a really hard ceiling on the complexity of what you can build. Uh, so I wanted to know where he got that number first. Uh, and I wanted to know if it was true. Uh, so you know the feeling that I'm talking about. You write this big beautiful skills file, pages of rules, edge cases, tone, formatting, you hand it to the agent, it does the thing. Uh, and you look at the output and go, did it actually pay attention? Did it actually follow all of these rules or did it just sort of, you know, do what it felt like and sort of give me a close simulacum of what I was expecting? Um, you can't really tell or can you? More on that later. Um, and so you live with this lowgrade anxiety every time you hit run. Um, and that feeling is what this research is about and what we're trying to find out if we can avoid. So here's my promise for your next 18 minutes. Uh, I'm going to show you where that 200 number came from, whether it's still true, and what the real number is today. Uh, because it moved by an order of magnitude. Uh, and then we're going to talk about what that means for you to take away. uh how long your skills and prompts can actually be and what that should what changes you should make to your workflow as a result. So the 200 number isn't folklore. Uh it comes from a real benchmark called IFScale uh from a paper uh by this guy whose name I'm going to mess up Jeroslowitch uh and co-authors last year. And the test is beautifully simple. Uh here's how if scale works. If you ask the model to write a business report uh and you give it a list of specific words that it has to include exactly in the report, include the exact word customer, include the exact word revenue, and so on for as many words as you want. Each of those is an instruction that it has to follow. Uh and then you count how many of those exact words showed up. Um so because the test is so simple, you only have to keep two numbers in your head. One is density, which we call n. That is how many rules we're talking about at once. And the second is accuracy, which is the percentage of those rules uh that it was able to actually follow. Uh now you might say uh that including random words in a report is not the same as following real instructions and fair enough and we're going to talk about that. Um but the keywords are a proxy. Uh include the word revenue is the same shape of task as include a section on pricing, right? Or never use this phrase. It is a discrete named constraint that you've told the agent that it has to follow. Um, if a model can't track 200 words in one prompt, it's definitely going to struggle with 200 more complicated instructions. Uh, so uh, if anything, it's going to do worse. So, this number is a ceiling. Uh, this number is as high as you can go. If you give it more complicated instructions, the number is probably going to get lower. And 200 is a really low ceiling. Um so before chasing new models you have to do good science which means that you have to replicate the uh old result and make sure uh that the 200 ceiling is real. So I reran the original benchmark. Um the original paper tested a whole batch of models uh and models uh live and die really fast. So uh by the time I got around to doing this testing only three of the models in the original set of 10 models that they used were still available via any kind of API. Uh so they were GPT 4.1, Claude Sonnet 4, and Gemini 2.5 Pro. Those were models that were available 12 months ago that are still available now. Um and that is why we tested those three because they were what was left. Um and since I first published this research a couple of weeks ago, uh one of those three models has been retired. So this was the last possible time that I could have run this test. Um so of that lineup, we're already down to two. So don't get attached to your models. Um here is the results that we got replicating the original if scale finding. Uh that is accuracy on the vertical axis. So it starts at 100% and begins to fall off. Uh and then the number of rules uh going up along the bottom on log scale. So every time it gets halfway across it has doubled uh the number of rules that it's dealing with. Um so by 500 rules you're losing 30 40 50% of them. Uh our curves matched the results in the original paper within the noise boundary. So the finding was real. uh a year ago somewhere around 200 to 300 rules frontier models started falling apart. That is a really low ceiling. Uh so that is our baseline and now comes the fun part where we took the exact same test and pointed it at the current frontier or rather what the current frontier was when I ran this test. So I ran GPT 5.5, Claude Opus 4.7 because 4.8 came out a week after I ran this test. Uh Gemini 3.1 Pro and Deepseek V4 Pro. So, I gave them the same prompt, the same words, the same everything. And I immediately ran into a problem, which is that they aced it. They all scored 100% immediately on this test. Absolutely no bugs. Uh, so we'd built a test to find the ceiling and the models had walked straight through the ceiling without noticing that the ceiling was there. Um, and that was a problem because the benchmark was written to top out at 500 words. So, I had to change the benchmark in order to be able to find the new ceiling. So, I moved the goalposts. I gave it more words to include. I doubled uh it from 500 to a,000. I doubled it again from a,000 to 2,000. And I kept doing that until I hit a 10,000word vocabulary. And that is where I began to find the ceiling of what models can do these days. Um, so let me put up the this is the money slide. This is the results. Remember log scale on the on the uh x-axis there. So it's going from 500 to 1,000 to 5,000 to 10,000. Uh so it looks like that scale is falling off of a cliff and it's actually happening over like a thousand numbers. Um but uh look how far to the right these new curves get before they bend. A year ago they were falling over at 200 to 300 instructions and now depending on the model the boundary is closer to 2,000. And for the best of them it is up to 5,000 instructions before they begin to fall off a cliff. So in about 12 months frontier models got close to 10 times better at following instructions simultaneously. That is the headline fighting and there is a lot of nuance that we need to get into. Um the capacity to track 2,000 named constraints in a single prompt is there. Um and that's really interesting because I think uh I don't know if everybody else feels this way but like it feel it felt to me like the the jump from you know GPT 5.1 to GPT 5.5 was kind of incremental, right? It didn't feel like we'd got 10 times better. But this is a test that really matters to uh a very practical thing like how long can my skills file be? Uh and in the course of a year we got 10 times better. Uh and that thing that gets me is that this benchmark is barely a a year old. A year later 500 is a rounding error. Uh and this keeps moving under my feet. I tested 4.7 uh opus 4.7. Opus 4.8 is even better. Um so this chart is a little out of date already which is kind of the whole point. If you set your engineering assumptions about how skills files should work, about how prompt how long your prompt can be, and you did that more than about six months ago, you are incorrect now, and you should probably be re-engineering how you do stuff. Uh, but there is more to this story uh because the way that the a models failed uh changed dramatically. Uh, and the way that they failed is very important. This part was a completely unexpected finding when I started running the experiment. Uh, and it totally messed up my test to start with because uh, the old failure mode was boring. They would just forget instructions and I could measure how many instructions they had remembered or forgotten. Uh, but the new ones fall apart in their own weird extremely onbrand way. Uh, so let me introduce you to how these four models fail. Uh, Deepseek 4 is a traditional model. It just forgets things. It doesn't have any drama. um it starts forgetting instructions around 750 rules and by 2000 it's dropping nearly half of them. Uh so it just forgets which frankly is the failure mode that I trust most because it's predictable. It's very easy to measure. Uh and the other models were not nearly as cooperative. Uh Opus 4.7 uh would decide repeatedly that the test was dangerous. Uh and what it would do is it would refuse at the API level to complete the test. I didn't know that there was an API response that you could get from Claude where it was like, "No, I could do this, but I'm not going to." Uh, but that's absolutely an API level response that Claude supports because they care so much about safety. Uh, and I started getting those all of the time. Uh, and the reason that was happening is because Claude has a very sensitive safety classifier. Uh, and if you put in certain combinations of words like say anthrax and cyanide, it decides that the whole request is dangerous and it bails out. Uh, and if you remember what my test does, my test is throwing uh 5 to 10,000 random words into uh into an instruction file. And so my my randomly selected words contained all sorts of things that looked dangerous in combination to the safety filter. And so it kept bailing saying that I was asking it to, you know, make a bomb or something. Um, so, uh, we had to for to get Claude to cooperate, I had to take all of my words and run them through OpenAI safety filter and filter out all of the naughty looking words so that it could get to anywhere. Once I given it that, Claude did really well. Uh so but the failure mode is that Claude is more likely to decide what you're doing is dangerous very early on uh at you know even two or 300 instructions if what you're doing uh is you know contains anything to do with medical advice because medical things often are dual purpose. They can be dangerous. They can be safe. Um so uh the third failure mode was Gemini 3.1 Pro. Gemini is rock solid all the way out uh to 5,000 instructions. It does extremely well. um genuinely one of the best on the chart. Uh and then past that it gets weird. Um it doesn't forget the instructions, it gets overwhelmed by the instructions. What it tries to do is it uh it uses thinking tokens to make sure that it is following all of the instructions at once. And when the number of instructions gets really high, it uses all of its thinking tokens. It uses its entire token budget thinking. And then it doesn't give any output. It's it gets to like nine, you know, if you've given it 10,000 tokens worth, it'll get 9,500 tokens worth of thinking and then give you a 500word response which doesn't contain any of the tokens. Uh so it thinks itself into a corner and runs out of room to actually answer, which is very expensive, uh and totally unhelpful, which is kind of on brand, isn't it? Um \[snorts\] uh which, you know, I would never say that out loud. Uh, and finally comes the winner, which is uh, GPT 5.5. GPT 5.5 is the best of the lot. 99% accuracy all the way out to 5,000 rules. Um, but if you push it far enough, it is by far the weirdest of the bunch. Uh, because it doesn't refuse outright. It doesn't silently forget. Instead, what it does is it gets frustrated and tells you that the test is stupid. Uh, it starts the report. It gets a few like that's the thing. It doesn't start out just saying no. It starts the report, it starts writing the report, and like 500 words into the report, it's like, "No, this is dumb. I'm not going to do this." And then it politely tells you, "This is dumb. I'm not going to do this anymore." Uh, that is the actual response that it gave me, but that that was like 5,000 words into the into this business report that I told it to generate. Um, so it's not wrong, right? I was asking for a coherent business report that on no particular subject that contains 5,000 random words. You're right, Gemini DPT. this is a a stupid thing to ask for. Um, which is a deeply unreasonable request and GPT called this out on it. Um, but it still counts as a failure in the test because the half-finish report that it gives you is missing most of the keywords and it is also the hardest one to detect because claude bails immediately. Claude says, "No, I'm not going to do this." Uh, Deepseek does its best. Uh, but GPT does what looks like a good job unless you read all the way to the end of the report where it says, "No, actually I'm going to bail because this is stupid." Um, so if you step back and look at the four together, Deep Sea quietly forgets, Claude gets scared and refuses, Gemini overthinks itself into silence, and GPT 5.5 finishes half of the job and tells you that the rest of it is beneath it. Um, and the point was the point isn't which one of these is funniest, although it is genuinely a little funny. Uh the point is that did it follow my instructions no longer has one failure mode. It has four different ways that it can fail and you can't recognize that failure unless you know which model you're dealing with and what its m what its pattern of failure is going to be. Uh so the models get 10 got 10x better. They fail in funny ways. Why should you care when you uh get back to your desk? Because three things have changed to your workflow. The first is that a year ago, the smart move was to keep every skills file very very short. Uh under 200 instructions, then point off to subsklls and a whole like you know byzantine labyrinth of uh additional skills files and subfiles and things like that. Uh and you mo you were compressing your your instructions to fit into a very small available space and you don't need to do that anymore. Your skills files can be very long. Um, number two is that if your use case needs a 100 specific rules or 300, you can just put them all in the prompt. Uh, you don't have to lie awake wondering whether which ones the model silently ignored. Um, and if you've been thinking uh about uh your own lived experience of using models, uh you probably recognize this. you've discovered that you've got less worried about how long your your prompt is going to get uh because the models have genuinely got 10 times better at following your prompts. Um 2,000 named constraints is an entire style guide, right? Like it's it's every brand rule, every legal disclaimer. Uh a year ago, you'd have had to shard that across a dozen specialized agents and hope that your specialized agents are hand are are handing off to each each other cleanly. But now you can ignore that. Um but the third thing is the big one. The question used to be can the model even do this? And the answer is now firmly yes. Well reasonably firmly. Uh is it worth the cost is the new question because you can include 10,000 words of of sorry 10,000 different instructions into your prompt. But that is going to be an enormous prompt. It's going to be a very expensive prompt. It's going to be a very slow prompt. So what used to be a hard wall that you would run against has now become a soft trade-off of is it worth me adding all of these extra instructions if it's going to give me more cost and more latency. Uh and now some caveats uh to head off the Q&A. Um first and important first and most important I mentioned this earlier this is a proxy task including random words uh in a in a fake business report um is evidence that long skills file works. It is not the same as proof that a long skills file works. Um, also the models hit the wall at wildly different points anywhere from 750 to 9,000 plus. So you have to pick your model very carefully. Uh, what our test doesn't do is measure whether the model reasoned clearly over a giant prompt. So uh, the good news is since I did my research several weeks ago, uh, a whole bunch of people have piled in on this. Um and now there's good research uh actual scientists have got involved and done uh Chroma's has done context rot work uh across 18 models showing that accuracy on long inputs can fall 30 to 50% well before you hit the context window limit. Uh and the weird part of their finding was that uh coherent well ststructured text is more likely to hit that failure mode uh than if you just put your instructions into a random order and shuffle them in. Uh, I don't know why that's the case. I'd have to read their report. Um, so the model can track 2,000, 5,000, possibly 10,000 instructions, but it's not necessarily going to uh reason clearly over them. It's not necessarily if those if those instructions conflict, if there is tension between them, it's not necessarily going to get that right. Um, and then there's the other one I mentioned briefly. Uh, collude's refusals are annoying, but they are loud. You get an error, you know it failed. Uh GPT's polite half-finish report is much more dangerous because it looks like a real answer. Uh you have to read the whole thing to notice that it gave up quietly halfway, which means that you can't trust the output. It means you have to read the output every single time to make sure whether or not it's working. Uh so the model will accept your 2,00 rules and it will hand you back something that looks at least to begin with confident and polished but could be bailing out halfway through. Um, so, uh, as an aside, people always ask me, "How much did all this cost me?" It cost me $29 to run all of these queries. 2,37 2,300 calls across seven models, uh, came to $29. Uh, it turns out novel research doesn't cost very much. Um, and this is the part of the talk where I was saying that you have to check this stuff in production because you can't trust that your model isn't going to silently fail. Uh, so you knew I was going to mention evals eventually because I work at Arise and this is where I do that. Um, but there are plenty of plugs for Arise. So I'm just going to say one true thing which is that if you are building a real AI application and you are giving it genuinely tricky tasks, you are going to run into one or more of these failure modes with a frontier model. Uh, and unless it's Claude telling you just to off at the API level, the only way to know that something went wrong is monitoring your outputs with another LLM. That is an eval. And that is what Arise does. And I'll leave it at that. Uh, I already mentioned that there's been new research since we did our own. Here's another important one. A paper landed testing 46 models called revisiting the reliability of language models in instruction falling, which you can bet made my ears perk up after I did that research myself. Uh, and they found something uncomfortable, which is that a model can ace a benchmark like ours and still be wildly unreliable. because if you reword the same instruction in a slightly different way, it can make a radical difference to how well uh it follows those instructions. So the model can follow 2,000 instructions and it can do it really well. But if you put the same instructions, the same 2,000 instructions in a different order, it can suddenly make the model much worse at following those instructions. And how ex how exactly to do that? what is the correct order of instructions to give your model such that it follows them perfectly as opposed to getting confused is still research that is being done. So capacity went up but reliability is still a problem. Um and then this is just a little brag because uh I was happy about it like I'm not a scientist. I did some research and then a whole bunch of other actual scientists piled in uh and did real science on the same question. There's now a whole bunch of benchmarks that have shown up uh to measure this same question. Firebench, CCR bench, Guidebench uh are all trying to measure the same thing. How well models follow a lot of real messy constraints at once. Uh and now the whole field is looking at it. So if you want better science than my, you know, 10,000 random words, uh the real science exists now. Uh so that gets me to where I will leave you. A year ago, the hard part of writing a skill was fitting everything in without the model losing the plot. That was a compression problem, and the compression problem is gone. uh the model will hold your 2,000 instructions just fine. The new hard part is knowing whether it actually did what you said and that is a verification problem. Uh a verification problem doesn't get solved by writing a better prompt. It gets solved by checking the output every time uh the same way that you would test any other code, which is to say an eval. The ceiling moved by 10x in one year. Uh so go back and check the assumptions that you made six months ago about how big your prompts should be, how big your uh instructions can get. uh because they might already be wrong. Boom. Be wrong. So that is the talk. If you want uh all of the code and all of the data, uh it is at this GitHub URL. Uh and this other QR code is uh something marketing made me insert. We are having a World Cup watch party tonight at 5:00 p.m. Uh you can come to our party. That link is to the Luma that will get you into the get into get you into the party. Uh I hope this talk has given you some novel information or at least a couple of laughs. And thank you so much for your time and attention.