# Five clues to what awaits us in artificial intelligence before the end of the year Author: Fernando Nieto Lobato Original publication: 2024-09-04 Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-49-avances-decisivos-en English URL: https://elcontemplador.github.io/estrategia-english/essays/049/ Status: Published translation This is a translation of the original Spanish essay published on 4 September 2024. Its claims, examples and forecasts retain that historical context. English publication: 2026-09-29 *Historical context: these expectations were published on 4 September 2024. Dates, prices, model names and predictions retain that perspective; this is not a current product or pricing guide.* After a summer that has been quieter in terms of major AI model releases, the arrival of autumn — and, above all, the passing of the crucial milestone of the US presidential election — seems to set the stage for a very eventful final four months of the year. We will see major advances that are sure to keep surprising us. Playing the futurist with AI, even over a very short period, is risky. [A few days ago, I read on Reddit about a discussion between two users predicting, in the comments on a post, when AIs might be able to generate an entire video game](https://www.reddit.com/r/singularity/comments/1f33rpy/they_said_50_yrs_i_said_5_yrs_it_literally_took_5/). One said we were still a long way off and it would take decades; the other thought it would take a few years. [It turned out that barely five days after that initial “discussion”, Google showed the world an AI system that fully simulated the classic game](https://hipertextual.com/2024/08/ia-generativa-gamengen-simular-doom-tiempo-real) *[Doom](https://hipertextual.com/2024/08/ia-generativa-gamengen-simular-doom-tiempo-real)*, [generated and run by AI](https://hipertextual.com/2024/08/ia-generativa-gamengen-simular-doom-tiempo-real). Predictions are almost impossible in this field. Even so, to leave you a little less unprepared, here are some clues to the most interesting developments that may arrive before the end of 2024. ## 1. Events The coming weeks and months are packed with AI events — OpenAI will hold an eagerly awaited developer day on 1 October — and general technology presentations in which AI will play a prominent part: Salesforce Dreamforce, Meta Connect, the launch of the new iPhones and so on. One announcement particularly caught our attention as confirmation that AI is becoming truly mainstream. On 29 August, it was announced that [Oprah Winfrey will host an ABC special on 12 September entitled “AI and the Future of Us”](https://www.forbes.com/sites/marcberman1/2024/08/29/oprah-winfrey-to-host-abc-special-ai-and-the-future-of-us/), featuring, among others, Bill Gates, Sam Altman and the director of the FBI. [![Historical video still of Oprah Winfrey, with a play symbol and the visible subtitle invent everything else.](https://elcontemplador.github.io/estrategia-english/assets/images/a30588a4-bfaa-4c2f-bd96-0f53136af9e4_610x1088.png)](https://substackcdn.com/image/fetch/$s_!a7IC!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa30588a4-bfaa-4c2f-bd96-0f53136af9e4_610x1088.png) *Original video still. The visible English subtitle is “invent everything else.” The preserved article contains this image; it is not an embedded playable video.* ## 2. The arrival of next-generation models By the end of this year, or early next year, we are highly likely to begin seeing a new generation of AIs, still based mainly on large language model architecture, offering a substantial qualitative leap over current models. To put it in more accessible terms, the expectation is that [the future GPT-5 — or Next, “Orion”, or whatever it is finally called, perhaps incorporating the mysterious Q* feedback technology](https://www.reddeperiodistas.com/openai-planea-integrar-su-proyecto-strawberry-en-chatgpt-en-otono/) — the next Claude model, [xAI's Grok 3, which Musk announced would be ready in December](https://www.binance.com/es-MX/square/post/2024-07-24-elon-musk-predicts-grok-3-0-to-be-most-powerful-ai-by-december-11232178206729), or even [Gemini 2, which could arrive in the next few weeks](https://x.com/OfficialLoganK/status/1829533022267625840), will show a meaningful advance over GPT-4-class models. Roughly speaking, they would move from the level of a secondary-school pupil or a student starting university to that of a PhD, improving above all in reasoning and avoiding hallucinations. The possible activation in some of these new AIs of what Kahneman called “System 2” — “thinking slowly” — through new machine-learning techniques used by models retraining themselves in increasingly effective ways, together with the benefits of scaling up, will probably give us AIs capable of handling more complex tasks reliably. It could also make them a dependable foundation for autonomous agents, something we are probably more likely to see next year. That would represent an enormous change in how we use AI, taking us far beyond chatbots or systems where we must supply detailed instructions, inspect the answer and manually repeat the process. ## 3. Nothing is real This is probably the issue with the most direct short-term consequences for politics. [If you have not already done so, I invite you to watch the following AI-generated video](https://x.com/thedorbrothers/status/1828813560878637165) — also our recommendation of the week — which is clearly satirical and exaggerated, and think for a few seconds about what it implies. [![Screenshot of the Dor Brothers' AI-generated satirical video post, with a synthetic depiction of Donald Trump; this is not footage of a real event.](https://elcontemplador.github.io/estrategia-english/assets/images/fae57e5c-bfc9-42be-9fa9-b53464aa9ae3_873x1146.png)](https://x.com/thedorbrothers/status/1828813560878637165) *Visible English post by The Dor Brothers, dated 28 August. The image is explicitly presented as AI-generated satire. The post is titled “The Hustle - Part 2” and reads:* > Experimenting with uncensored AI generators like @grok and Flux to create some unhinged, satirical content 🙌💯 @elonmusk > > Song: Hanumankind - Damnson > > Please Note: We don't take political stances or aim to spread fake news or fear. While we believe the *The disclaimer ends there in the screenshot; the remainder is hidden behind “Show more”.* By late 2024, video deepfakes and voice cloning really are approaching the critical point at which they become virtually indistinguishable from reality. More importantly, they are beginning to reach almost anyone with slightly advanced technical skills, before becoming an everyday technology we can use on our own phones within months or a few years. That video is one of the best examples. Although further refinement is needed, it is already entirely possible to clone someone's appearance and voice and create such montages very quickly, without extensive expertise. Surprising as it may seem, this has even reached **live** images and audio, as [this example of a fake Elon Musk broadcasting live with an “AI mask” and a voice-cloning system](https://www.tiktok.com/@brandnat/video/7407810794592963848) shows. It was thought that we would not see this sort of thing until at least the end of the year. That has been, and remains, true of major companies such as Google and OpenAI, which are holding back their best and most realistic image and video models until after the US election. But new entrants have made it a reality we will have to live with. The genie is out of the bottle. Even if regulation can be improved, we will have to get used to the vertiginous leap of living in a world where we can no longer simply believe that anything we see on a screen, however plausible it looks, is “real”. ## 4. AI prices are falling dramatically, making it easier for AI to be everywhere In less than 24 months, the cost of two million tokens of “top-of-the-range” AI has fallen from US$180 to 75 cents: 240 times cheaper. These and similar figures suggest that AI is becoming a commodity at a certain level, with a trend towards virtually free “artificial intelligence”, except for the latest frontier models. Against this backdrop of continuously falling prices, if capabilities keep improving — at a pace we have yet to see — AI can be expected to enter more and more human productive processes. This, in turn, will widen the productivity gap with people and their wages still further. It will undoubtedly have growing consequences for employment, and politicians should already be devoting a substantial share of their time to considering solutions. [![Historical posts by Ben EB and Andrew Ng discussing falling token costs, with the embedded Andrew Ng post cut off before its conclusion.](https://elcontemplador.github.io/estrategia-english/assets/images/58dd42fd-3cb3-4487-b140-a10c3e3b3a59_877x496.png)](https://substackcdn.com/image/fetch/$s_!VsNU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F58dd42fd-3cb3-4487-b140-a10c3e3b3a59_877x496.png) *Visible English screenshot text. Ben EB writes:* > At 79% cost reduction every 12 months, LLM token cost per million is decreasing faster than the pace of compute power under Moore’s Law (50% 24mths). > > “…even if an application is marginally too expensive to run today, it may be worth deploying in anticipation of lower prices.” *The embedded Andrew Ng post, dated 29 August, reads:* > After a recent price reduction by OpenAI, GPT-4o tokens now cost $4 per million tokens (using a blended rate that assumes 80% input and 20% output tokens). GPT-4 cost $36 per million tokens at its initial release in March 2023. This price reduction over 17 months corresponds to *The remainder is not visible. These are historical claims with the quantities and assumptions shown in the screenshot, not current price quotations.* We are very probably still in the earliest stages of productive AI adoption. But with this prospect of almost-free intelligence, adoption rates could rise very quickly in the near future. We are already seeing notable advances in programming, where models are progressing in leaps and bounds, and even in something AI has found harder than it seemed it would a decade ago: autonomous driving. [Waymo's self-driving taxis have gone from 10,000 paid trips a week to more than 100,000 in a year](https://www.reddit.com/r/singularity/comments/1ex29lx/waymo_has_surpassed_100k_paid_trips_per_week_it/?share_id=oNVBII9BBinPQCrWywRqf). ## 5. Technologies already promised and demonstrated reach consumers: interactive voice in ChatGPT, Sora, Veo… Our final clue to the year's last four months concerns highly promising technologies we have already seen but which have not yet reached the general public. That may be because of the serious risks of misuse — particularly before an event such as the US election — or because serving many millions of users at full capacity requires enormous computing resources. Especially prominent here are ChatGPT's new voice model and Sora, the OpenAI video model that astonished the world earlier this year. A small group is already testing the former in a private beta, with a general release promised before the end of autumn. A post-election release date for Sora would not surprise us at all: if OpenAI delays too long, the video field is attracting a great many new competitors with increasingly powerful products. All in all, the next four months look highly interesting for AI and its “political” implications: its popularisation through programmes such as Oprah's; the role increasingly convincing and easy-to-create AI deepfakes might play in the major US election campaign, and subsequently elsewhere; and the enormous competitive advantage a country, organisation or candidate might gain by having access to a new, state-of-the-art language model before its rivals. And without doubt, one of the most exciting things about this field is the strong likelihood of another major surprise along the way. Fernando Nieto Lobato *Director of Digital Innovation, Institución Educativa ALEPH*