GPT-4 PDF Upload Resumes: Coding as a Career

From uploading a PDF to rewriting employment: ChatGPT-4 opens another league

You upload a PDF. Or an entire folder. The tool returns the summary in the lines you ask for. Those who have tested it summarize it this way: the latest version «plays in a different league». The detail that set everything off is minimal: native document upload stopped giving errors on the afternoon of the 10th, and until then it only accepted standalone image formats. From then on, the conversation stopped being about chatbots and started being about payrolls and what is worth studying.

How much does it cost and what changes with the new version?

24 euros a month. That is the price mentioned to access the latest version, and those who pay for it describe it as opening «a world of possibilities». Before the update, uploading a document was a crucifixion: the tool rejected files and forced users to use plugins with a page limit. Now it swallows PDFs and folders.

Image generation is the other leap highlighted. Several accounts insist that the result exceeds the user's previous idea, to the point that someone claims to have obtained better sketches than those they had in mind for their own business project. Custom digital character creation appears in the same sentence.

Those who want to squeeze it out will have to put in their part. Performance depends on the instructions written, and those instructions are only good if the subject is mastered. Those who do not master it delegate to the machine itself and get correct but flat results.

The IBEX department already testing its own cut

The most concrete case is that of an administrative department of a large IBEX company. There, a project has been set up with ChatGPT that, according to the shared calculation, would eliminate 15% of the staff. These are profiles that process files, fill out forms, and read and write email. The tests, say those who have done them, come out «acceptably well».

The nuance matters: the system's error rate resembles the human errors that already occur. The comparison is not against perfection, but against the current staff. That is the argument that weighs most in the pessimistic part of the matter.

From 10 lawyers to 4, and from 100 judges to 20

In the legal sector, the same pattern is described. A large firm that previously assigned ten lawyers to a medium-sized province, six of them interns searching for case law, could be left with four. The part of the work that goes is exactly what was never segarro.

In justice, the projection is more aggressive. With so-called predictive justice drafting judgments that the human judge would only review, the circulating calculation reduces one hundred judges to twenty. It is advisable to take it for what it is: an estimate by the writer, not an approved plan.

From there comes the macroeconomic corollary that some extract without holding back: reduced working hours, universal basic income, and a money printer as the only way out. It is said half in jest. It is said.

Teachers counterattack in the classroom

Students already use it to do exercises. And part of the teaching staff assumes it and changes strategy: if the machine solves the standard prompt, the teacher becomes more creative and poses problems that cannot be solved without thinking. Some tell it with great pleasure, with the teacher himself asking the AI to fabricate exercises impossible for the AI to solve.

The education system comes out badly in several messages, and not just due to the technological intrusion. It is maintained that the traditional exam format disadvantages the more introverted students and that a free evaluation model would suit them better.

The end of code pecking: what happens to software engineering?

Here the discussion sharpens, because not all programmers are the same. A system administrator explains it with a repeated analogy: the one who «pecks code» is to a software engineer as the one who lays bricks is to an architect. From this perspective, what is automated first is the mechanical part.

The sector itself admits that much commercial software is manufactured with time and budget criteria before quality. If code could be generated from a base of tested functions, the savings would be immediate. It is also said that a GPT configured as a Java editor got a 10 in the programming course of a training cycle.

And there is a rather elegant underlying argument: prompts are a programming language in natural language. If this holds, the blow is not that the machine writes Java or Python, but that the end user no longer needs intermediaries to ask for what they want.

What AI still does not solve: data, mathematics, and systems

It is not all estimulante ilegal. An analysis points to the real bottleneck: data quality. AI has been fed by what people gave away on social media and search engines, and when it starts eating its own production, the risk of degradation is real. That was, in fact, the answer the machine itself gave when asked about its brakes.

It is also remembered that there is field beyond programming websites and APIs: machine learning, computer vision, graphics. And that there, mathematics is not optional, no matter how the sector sells itself as a quick exit. Another note distinguishes between administering Unix servers, a quiet life, and fighting with Windows, which is described as a tightrope walk.

And a final, uncomfortable piece of data: some maintain that version 4 programs worse than version 3. With that on the table, the question of what to study has no single answer, and the only certainty is that the list of «safe» jobs is shortening.