Welcome to our annual report on the utilization of the O’Reilly studying platform. It’s been an thrilling yr, dominated by a relentless stream of breakthroughs and bulletins in AI, and sophisticated by industry-wide layoffs. Generative AI will get higher and higher—however that development could also be at an finish. Now the ball is within the utility builders’ courtroom: Where, when, and the way will AI be built-in into the purposes we construct and use on daily basis? And if AI replaces the builders, who will likely be left to do the mixing? Our information reveals how our customers are reacting to adjustments within the {industry}: Which expertise do they should brush up on? Which do they should add? What do they should know to do their day-to-day work? In brief: Where have we been up to now yr, and the place are we going?
We aren’t involved about AI taking away software program builders’ jobs. Ever for the reason that laptop {industry} received began within the Nineteen Fifties, software program builders have constructed instruments to assist them write software program. AI is simply one other software, one other hyperlink added to the top of that chain. Software builders are excited by instruments like GitHub Copilot, Cursor, and different coding assistants that make them extra productive.
Learn sooner. Dig deeper. See farther.
That’s solely one of many tales we’re following. Here are just a few of the others:
- The subsequent wave of AI improvement will likely be constructing brokers: software program that may plan and execute advanced actions.
- There appears to be much less curiosity in studying about programming languages, Rust being a major exception. Is that as a result of our customers are keen to let AI “learn” the main points of languages and libraries for them? That is perhaps a profession mistake.
- Security is lastly being taken critically. CEOs are uninterested in being within the information for the improper causes. AI instruments are beginning to take the load off of safety specialists, serving to them to get out of “firefighting” mode.
- “The cloud” has reached saturation, not less than as a talent our customers are learning. We don’t see a surge in “repatriation,” although there’s a fixed ebb and movement of knowledge and purposes to and from cloud suppliers.
- Professional improvement may be very a lot of curiosity to our customers. Specifically, they’re targeted on being higher communicators and main engineering groups.
All of those traits have been impacted, if not pushed, by AI—and that impression will proceed within the coming yr.
Finally, some notes about methodology. Skip this paragraph in order for you; we don’t thoughts. This report is predicated on the usage of O’Reilly’s on-line studying platform from January 1, 2024, to September 30, 2024. Year-over-year comparisons are based mostly on the identical interval in 2023. The information in every graph is predicated on O’Reilly’s “units viewed” metric, which measures the precise use of every merchandise on the platform. It accounts for various utilization conduct for various media: textual content, programs, and quizzes. In every graph, the information is scaled in order that the merchandise with the best models considered is 1. That means gadgets inside a graph are comparable to one another, however you may’t examine an merchandise in a single graph to an merchandise in one other. And all percentages are reported with two important digits.
Skills
When we take a look at how our clients use the O’Reilly studying platform, we all the time assume by way of expertise. What expertise are they attempting to realize? And how are they attempting to enhance their data? This yr, one thread that we see throughout all of our platform is the significance of synthetic intelligence. It’s all about upskilling within the age of AI.
Artificial Intelligence
It will shock completely no one that AI was essentially the most lively class up to now yr. For the previous two years, massive fashions have dominated the information. That development began with ChatGPT and its descendants, most just lately GPT 4o1. But not like 2022, when ChatGPT was the one present anybody cared about, we now have many contenders. Claude has emerged as a favourite amongst programmers. After a shaky begin, Google’s Gemini fashions have grow to be strong performers. Llama has established itself as one of many high fashions and because the matriarch of a wealthy ecosystem of open1 fashions. Many of the open fashions can ship acceptable efficiency when operating on laptops and telephones; some are even focused at embedded gadgets.
So what does our information present? First, curiosity in nearly all the high expertise is up: From 2023 to 2024, Machine Learning grew 9.2%; Artificial Intelligence grew 190%; Natural Language Processing grew 39%; Generative AI grew 289%; AI Principles grew 386%; and Prompt Engineering grew 456%. Among the highest matters, essentially the most important decline was for GPT itself, which dropped by 13%—not an enormous decline however definitely a major one. Searches for GPT peaked in March 2023 and have been trending downward ever since, so our search information matches our utilization information.
We’re used to seeing curiosity transfer from a extra common high-level subject to particular expertise as an {industry} sector matures, so this development away from GPT in favor of extra summary, high-level matters is counterintuitive. But in context, it’s pretty clear what occurred. For all sensible functions, GPT was the one sport on the town again in 2023. The scenario is totally different now: There’s numerous competitors. These different fashions don’t but present up considerably in search or utilization information, however the customers of our platform have discovered what’s necessary: not studying about GPT or Claude or Gemini or Mistral however getting the background you should make sense of any mannequin. Discovering a workflow that matches your wants is necessary, and as Simon Willison factors out, your ultimate workflow may very well contain utilizing a number of fashions. Recent fashions are all good, however they aren’t all good in the identical manner.
AI has had a fantastic yr, however will it proceed to point out features in 2025? Or will it drop again, a lot as ChatGPT and GPT did? That is dependent upon many components. Gartner has generative AI slipping into the “trough of disillusionment”—and no matter you consider the expertise’s promise, keep in mind that the disillusionment is a sociological phenomenon, not a technical one, and that it occurs as a result of new applied sciences are overhyped. Regardless of generative AI’s long-term promise, we anticipate some disillusionment to set in, particularly amongst those that haven’t correctly understood the expertise or its capabilities.
Prompt Engineering, which gained 456% from 2023 to 2024, stands out. A 456% achieve isn’t as shocking because it appears; in any case, folks solely began speaking about immediate engineering in 2023. Although “prompt engineering” was bandied about as a buzzword, it didn’t grow to be a talent that employers have been on the lookout for till late in 2023, if that. That could also be an early warning sign for AI disillusionment. Searches for “prompt engineering” grew sharply in 2023 however appeared to say no barely in 2024. Is that noise or sign? If disillusionment in Prompt Engineering units in, we’ll additionally see declines in higher-level matters like Machine Learning and Artificial Intelligence.
There’s a special tackle the way forward for immediate engineering. There have been a variety of arguments that the necessity for immediate engineering is short-term. As generative AI improves, this line of reasoning contends, we are going to not want to jot down advanced prompts that specify precisely what we wish the AI to do and do it. Prompts will likely be much less delicate to precisely how they’re worded; altering a phrase or two will not give a totally totally different outcome. We’ll not should say “explain it to me as if I were five years old” or present a number of examples of remedy an issue step-by-step.
Some current developments level in that course. Several of the extra superior fashions have made the “explain it to me” prompts superfluous. OpenAI’s GPT 4o1 has been educated in a manner that maximizes its problem-solving talents, not simply its skill to string collectively coherent phrases. At its greatest, it eliminates the necessity to write prompts that show remedy the issue (a method referred to as few-shot prompting). At worst, it “decides” on an inappropriate course of, and it’s tough to persuade it to unravel the issue a special manner. Anthropic’s Claude has a brand new (beta) laptop use function that lets the mannequin use browsers, shells, and different applications: It can click on on hyperlinks and buttons, choose textual content, and do far more. (Google and OpenAI are reportedly engaged on related options.) Enabling a mannequin to make use of the pc in a lot the identical manner as a human seems to present it the flexibility to unravel multistep issues by itself, with minimal description. It’s an enormous step towards a future stuffed with clever brokers: linked AI programs that cooperate to unravel advanced issues. However, Anthropic’s documentation is filled with warnings about critical safety vulnerabilities that stay to be solved. We’re thrilled that Anthropic has been forthright about these weaknesses. But nonetheless, whereas laptop use could also be a peek on the future, it’s not prepared for prime time.
AI will nearly definitely slide right into a trough of disillusionment; as I’ve mentioned, the trough has extra to do with sociology than with expertise. But OpenAI and Anthropic are demonstrating necessary paths ahead. Will these experiments bear fruit within the subsequent yr? We’ll see.
Many expertise related to AI additionally confirmed strong features. Use of content material about Deep Learning is up 14%, Generative Models is up 26%, and GitHub Copilot is up 471%. Use of content material in regards to the main AI libraries was up barely: PyTorch gained 6.9%, Keras elevated 3.3%, and Scikit-Learn gained 1.7%. Usage of TensorFlow content material declined 28%; its continued decline signifies that PyTorch has gained the hearts and minds of AI builders.
These features—notably Copilot’s—are spectacular, however a extra necessary story issues two expertise that got here out of nowhere: Usage of content material about LangChain is on a par with PyTorch, and RAG is on a par with Keras. Neither of those expertise have been in final yr’s report; in 2023, content material utilization for LangChain and RAG was minimal, largely as a result of little content material existed. They’ve caught on as a result of each LangChain and RAG are instruments for constructing higher purposes on high of AI fashions. GPT, Claude, Gemini, and Llama aren’t the top of the highway. RAG allows you to construct purposes that ship personal information to a mannequin as a part of the immediate, enabling the mannequin to construct solutions from information that wasn’t in its coaching set. This course of has a number of necessary penalties: It minimizes the likelihood of error or “hallucination”; it makes it potential to attribute solutions to the sources from which they got here; and it typically makes it potential to make use of a a lot smaller and extra economical mannequin.
LangChain is the primary of many frameworks for constructing AI brokers. (OpenAI has Swarm; Google has an Agent Builder that’s a part of Vertex; Salesforce and different distributors even have choices.) Agents are software program that may plan and execute multistage actions, lots of that are delegated to different AI fashions. Claude’s laptop use API is one other side of this development, together with no matter merchandise OpenAI and Google could also be constructing. Saying that utilization has elevated 26 million % isn’t to the purpose—however realizing that LangChain has grown from close to zero to a platform on a par with PyTorch may be very a lot so. Agentic purposes are definitely the subsequent large development inside AI.
Data
Artificial intelligence depends closely on what we used to name (and maybe nonetheless name) information science. Building AI fashions requires information at unprecedented scale. Building purposes with RAG requires a portfolio of knowledge (firm financials, buyer information, information bought from different sources) that can be utilized to construct queries, and information scientists know work with information at scale.
Therefore, it’s not shocking that Data Engineering expertise confirmed a strong 29% enhance from 2023 to 2024. SQL, the frequent language of all database work, is up 3.2%; Power BI was up 3.0%, together with the extra common (and far smaller) subject Business Intelligence (up 5.0%). PostgreSQL is near edging forward of MySQL, with a 3.6% achieve. Interest in Data Lake architectures rose 59%, whereas the a lot older Data Warehouse held regular, with a 0.3% decline. (In our talent taxonomy, Data Lake contains Data Lakehouse, a knowledge storage structure that mixes options of knowledge lakes and information warehouses.) Finally, ETL grew 102%. With the exception of ETL, the features are smaller than the will increase we noticed for AI expertise, however that is sensible: AI is an thrilling new space, and information is a mature, steady class. The quantity of people that want specialised expertise like ETL is comparatively small however clearly rising as information storage turns into much more necessary with AI.
It’s price understanding the connection between information engineering, information lakes, and information lakehouses. Data engineers construct the infrastructure to gather, retailer, and analyze information. The information wanted for an AI utility nearly all the time takes many varieties: free-form textual content, photos, audio, structured information (for instance, monetary statements), and many others. Data typically arrives in streams, asynchronously and kind of continuously. This is an effective match for a knowledge lake, which shops information no matter construction to be used later. Because information receives solely minimal processing when it arrives, it may be saved in close to actual time; it’s cleaned and formatted in application-specific methods when it’s wanted. Once information has been saved in a knowledge lake, it may be used for conventional enterprise analytics, saved in a vector or graph database for RAG, or put to nearly every other use. An information lakehouse combines each structured and unstructured information in a single platform.
Software Development
What do software program builders do all day? They write software program. Programming is a vital a part of the job, nevertheless it’s not the entire thing; greatest estimates are that programmers spend roughly 20% of their time writing code. The remainder of their time is spent understanding the issues they’re being requested to unravel, designing acceptable options, documenting their work, updating administration on the standing of their tasks, and far more.
Software structure, which focuses on understanding a buyer’s necessities and designing programs to fulfill these necessities, is a vital a part of the general software program improvement image. It’s a talent to which lots of our software program builders and programmers aspire.
Architecture
This yr’s information reveals that software program structure continues to be one of the crucial fascinating expertise within the industries we serve. Usage of fabric about Software Architecture rose 5.5% from 2023 to 2024, a small however important enhance. But it’s extra necessary to ask why it elevated. A place in software program structure could also be perceived as safer in a time of layoffs, and it’s typically perceived as one other step ahead in a profession that strikes from junior programmer to senior to steer. In addition, the rise of AI presents many architectural challenges: Do we attempt to construct our personal mannequin? (The reply is often “no.”) Should we use an AI service supplier like OpenAI, Anthropic, Microsoft, or Google, or ought to we fine-tune and host our personal mannequin on our personal infrastructure? How will we construct purposes which might be secure (and the way will we outline “safe”)? How will we consider efficiency? These questions all have a bearing on software program structure. Furthermore, AI would possibly present instruments to assist software program architects, however up to now, these instruments can do little for the substance of the job: understanding clients’ wants and serving to them outline what they wish to construct. With AI within the image, we’re all constructing new sorts of purposes—and people purposes require architects to assist design them.
In this context, it’s no shock that Enterprise Architecture is up 17% and Distributed Systems is up 35%. Enterprise structure is a staple: As Willie Sutton mentioned about banks, “That’s where the money is.” It’s a great wager that many enterprises are attempting to combine AI into their programs or replace legacy programs which might be not scalable or maintainable. We can (and do) make the identical argument about distributed programs. Modern enterprises work on a scale that was unimaginable just a few many years in the past. Scale isn’t only for firms like Amazon and Google. To survive, even small companies must develop a web based presence—and meaning constructing programs within the cloud that may deal with surges in demand gracefully. It means constructing programs that may stand up to outages. Distributed programs aren’t simply large deployments with a whole lot of hundreds of nodes. Your enterprise might solely require a dozen nodes, however whatever the scale, it nonetheless faces the architectural challenges that include distributed programs.
Some of the extra important concepts from the previous decade appear to be falling out of favor. Microservices declined 24%, although content material use remains to be substantial. Domain-Driven Design, which is a superb talent for designing with microservices, is down 22%. Serverless is down 5%; this specific architectural fashion was broadly hyped and appeared like a great match for microservices however by no means actually caught on, not less than based mostly on our platform’s information.
What’s occurring? Microservice architectures are tough to design and implement, and so they aren’t all the time acceptable—from the beginning, the very best recommendation has been to start by constructing a monolith, then break the monolith into microservices when it turns into unwieldy. By the time you attain that stage, you’ll have a greater really feel for what microservices have to be damaged out from the monolith. That’s good recommendation, however the hype received forward of it. Many organizations that will by no means want the complexity of microservices have been attempting to implement them with underskilled employees. As an architectural fashion, microservices gained’t disappear, however they’re not getting the eye they as soon as have been. And new concepts, like modular monoliths, might catch on within the coming years; modularity is a advantage no matter scale or complexity.
Programming languages
Last yr’s report confirmed that our customers have been consuming much less content material about programming languages. This yr’s information continues that development. We see a small drop for Python (5.3%) and a extra important drop for Java (13%). And even C++, which confirmed wholesome progress from 2022 to 2023, is down 9% in 2024.
On the opposite hand, C is up (1.3%), and so is C# (2.1%). Rust is up 9.6%. The small will increase in C and C# could be noise. C is well-entrenched and isn’t going anyplace quick. Neither is C++, regardless of its drop. Rust’s enhance continues a progress development that stretches again a number of years; that’s an necessary sign. Rust is clearly successful over builders, not less than for brand new tasks. Now that the US authorities is putting a precedence on reminiscence security, Rust’s emphasis on reminiscence security serves it effectively. Rust isn’t the primary programming language to assert reminiscence security, nor will or not it’s the final. (There are tasks to add reminiscence security to C++, for instance.) But proper now, it’s the very best positioned.
Aside from Rust, although, we have to ask what’s occurring with programming expertise. A number of forces are making use of downward strain. Industry-wide layoffs could also be enjoying a job. We’ve downplayed the impact of layoffs up to now, however we might should admit that we have been improper: This yr, they could be taking a chew out of expertise improvement.
Could generative AI have had an impact on the event of programming language expertise? It’s potential; shortly after GPT-3 was launched, Simon Willison reported that he was studying Rust with the assistance of ChatGPT and Copilot, and extra just lately that he’s used Claude to write Rust code that he has in manufacturing, regardless that he doesn’t think about himself a talented Rust developer.
It can be silly to disclaim that generative AI will assist programmers to grow to be extra productive. And it will be silly to disclaim that AI will change how and what we study. But now we have to consider carefully about what “learning” means, and why we study within the first place. Programmers gained’t have to recollect all of the little particulars of programming languages—however that’s by no means been the necessary a part of programming, nor has rote memorization been an necessary a part of studying. Students won’t ever have to recollect a half dozen sorting algorithms, however laptop science lessons don’t educate sorting algorithms as a result of committing algorithms to reminiscence is necessary. Every programming language has a form() operate someplace in its libraries. No, sorting is taught as a result of it’s an issue that everybody can perceive and that may be solved in a number of other ways—and every resolution has totally different properties (efficiency, reminiscence use, and many others.). The level is studying remedy issues and understanding the properties of these options. As Claire Vo mentioned in her episode of Generative AI within the Real World, we’ll all the time want engineers who assume like engineers—and that’s what studying remedy issues means. Whether strains finish in a semicolon or a colon or whether or not you employ curly braces, finish statements, or tabs to delimit blocks of code is immaterial.
The notion that generative AI minimizes the necessity to study programming languages might restrict the usage of language-oriented content material on our platform. Does that profit the learners? If somebody is utilizing AI to keep away from studying the exhausting ideas—like fixing an issue by dividing it into smaller items (like quicksort)—they’re shortchanging themselves. Shortcuts hardly ever repay in the long run; coding assistants might enable you to to jot down some helpful code, however those that use them merely as shortcuts moderately than as studying instruments are lacking the purpose. Unfortunately, the historical past of instructing—going again centuries if not millennia—has confused memorization. It’s time for each learners and lecturers to develop past that.
Learning is altering because of AI. The manner we educate, and the way in which our customers wish to be taught, is altering. Building the correct of experiences to facilitate studying in an AI-enabled atmosphere is an ongoing venture for our studying platform. In the long run, will our customers study to program by finishing AI-generated tutorials which might be custom-made in actual time to their wants and skills? That’s the place we’re headed.
Web programming
Use of content material about net programming expertise is down, with few exceptions. Quite a few components is perhaps contributing to this. First, I can’t consider any important new net frameworks up to now yr; the sector remains to be dominated by React (down 18%) and Angular (down 10%). There is a few life close to the underside of the chart. The Svelte framework had important progress (24%); so did Next.js (8.7%). But whereas these frameworks have their adherents, they’re removed from dominant.
PHP (down 19%) nonetheless claims to have constructed the lion’s share of the net, nevertheless it’s not what builders attain for after they wish to construct one thing new, notably if that “new” is a fancy net utility. The PHP world has been rocked by a bitter battle between the CEOs of Automattic (the builders of WordPress, by far crucial PHP framework) and WP Engine (a WordPress internet hosting platform). That battle began too late to have an effect on this yr’s outcomes considerably, nevertheless it would possibly weigh closely subsequent yr.
A extra important improvement has been the motion away from advanced platforms and again towards the simplicity of the sooner net. Alex Russell’s “Reckoning” posts summarize most of the issues. Our networks and our computer systems are a lot, a lot sooner than they have been 20 or 25 years in the past, however net efficiency hasn’t improved noticeably. If something, it’s gotten worse. We nonetheless await purposes to load. Applications are exhausting to develop and have gotten tougher over time. There are a number of new frameworks which will (or might not) be lighter-weight, akin to HTMX, Ludic, Glitch, and Cobalt. None of them have but made a dent in our information, partially as a result of none have constructed sufficient of a following for publishers and trainers to develop content material—and you may’t have any models considered if there isn’t something to view. However, in order for you an expertise that isn’t dominated by heavyweight frameworks, doesn’t require you to grow to be a JavaScript professional, and places the enjoyable again into constructing the net, that is the place to look.
Web dev is a self-discipline that has been ill-served by shortcuts to studying. We hear too typically about boot camp graduates who know just a few React tips however don’t perceive the distinction between React and JavaScript (and even know that JavaScript exists, not to mention different programming languages). These programmers are very more likely to lose their jobs to AI, which may already reproduce all the essential React strategies they’ve realized. Learning suppliers want to consider how AI is altering the office and the way their college students can companion with AI to construct one thing past what AI can construct by itself. Part of the answer is definitely a return to fundamentals, making certain that junior builders perceive the instruments with which they’re working.
IT Operations
Operations is one other space the place the traits are largely downward. It could also be small comfort, however the drops for a number of of crucial matters are comparatively small: Linux is down 1.6%, Terraform is down 4.0%, and Infrastructure as Code is down 7.3%. As a talent, Terraform appears little harm by the fork of Terraform that created the open supply OpenTofu venture, maybe as a result of the OpenTofu builders have been cautious to take care of compatibility with Terraform. How this cut up performs out sooner or later is an open query. It’s price noting the precipitous drop in Terraform certification (down 43%); which may be a extra necessary sign than Terraform itself.
Kubernetes is down 20%. Despite that drop, which is sharper than final yr’s 6.9% lower, content material instructing Kubernetes expertise stays the second most generally used group on this class, and Kubernetes certification is up 6.3%. Last yr, we mentioned that Kubernetes wanted to be less complicated. It isn’t. There are not any viable alternate options to Kubernetes but, however there are other ways to deploy it. Kubernetes as a service managed by a cloud supplier is definitely catching on, placing the burden of understanding each element of Kubernetes’s operation on the shoulders of the supplier. We additionally pointed to the rise of developer platforms; this yr, the buzzword is “platform engineering” (Camille Fournier and Ian Nowland’s ebook is superb), however so far as Kubernetes is worried, it’s the identical factor. Platform engineers can summary data of Kubernetes right into a platform, minimizing software program builders’ cognitive overhead. The result’s that the quantity of people that must find out about Kubernetes is smaller.
Both DevOps (down 23%) and SRE (down 15%) dropped. There’s definitely some frustration with DevOps: Has it paid off? We ask a special query: Has it ever been tried? One drawback with DevOps (which it shares with Agile) is that many firms “adopted” it in title however not in essence. They renamed just a few positions, employed just a few DevOps engineers, perhaps created a DevOps group, by no means realizing that DevOps wasn’t about new job titles or new specialties; it was about lowering the friction between software program improvement groups and operations groups. When you take a look at it this fashion, creating new teams and hiring new specialists can solely be counterproductive. And the result’s predictable: You don’t should look far to search out blogs and whitepapers claiming that DevOps doesn’t work. There’s additionally frustration with concepts like “shift left” and DevSecOps, which envision taking safety under consideration from the beginning of the event course of. Security is a special dialogue, nevertheless it’s unclear the way you construct safe programs with out taking it under consideration from the beginning. We’ve spent a number of many years constructing software program and attempting to fold safety in on the final minute—we all know how effectively that works.
In any case, the {industry} has moved on. Platform engineering is, in some ways, a pure outgrowth of each DevOps and SRE. As I’ve argued, the course of operations has been to extend the ratio of computer systems to operators. Is platform engineering the subsequent step, permitting software program builders to construct programs that may deal with their very own deployment and routine operations with out the assistance of operations employees?
IT certifications
General IT certifications, aside from safety, trended downward. Use of content material to organize for the CompTIA A+ examination, an entry-level IT certification, was down 15%; CompTIA Network+ was down 7.9%. CompTIA’s Linux+ examination held its personal, with a decline of 0.3%. On our platform, we’ve seen that Linux sources are in excessive demand. The slight decline for Linux-related content material (1.6%) matches with the very small lower in Linux+ certification.
For a few years, Cisco’s certifications have been the gold customary for IT. Cisco Certified Network Associate (CCNA), a reasonably common entry-level IT certification, confirmed the best utilization and the smallest decline (2.2%). Usage of content material to organize for the Cisco Certified Network Practitioner (CCNP) exams, a cluster of associated certifications on matters like enterprise networking, information facilities, and safety, dropped 17%. The Cisco Certified Internet Engineer (CCIE) exams confirmed the best decline (36%). CCIE has lengthy been acknowledged as essentially the most complete and in-depth IT certification. We’re not stunned that the overall utilization of this content material is comparatively small. CCIE represents the climax of a profession, not the beginning. The quantity of people that attain it’s comparatively small, and those that do typically embody their CCIE quantity with their credentials. But the drop is shocking. It’s definitely true that IT is much less targeted on heavy-duty routing and switching for on-prem information facilities (and even smaller machine rooms) than it was just a few years in the past. That work has largely been offloaded to cloud suppliers. While routers and switches haven’t disappeared, IT doesn’t must assist as large a variety of sources: They must assist workplace WiFi, some databases that want to stay on-premises, and perhaps just a few servers for office-related duties. They’re very involved about safety, and as we’ll see shortly, safety certifications are thriving. Is it potential that Cisco and its certifications aren’t as related as they was once?
As we talked about above, we additionally noticed a drop within the comparatively new certification for HashiCorp’s Terraform (43%). That’s a pointy decline—notably since use of content material about Terraform itself solely declined 4.0%, displaying that Terraform expertise stay extremely fascinating whatever the certification. A sudden drop in certification prep may be brought on by a brand new examination, making older content material out-of-date, however that isn’t the case right here. Terraform certification definitely wasn’t helped by HashiCorp’s swap to a Business Source License or the following fork of the Terraform venture. IBM’s pending acquisition of Terraform (set to shut earlier than the top of 2024) might have launched extra uncertainty. Is the decline in curiosity for Terraform certification an indicator of dissatisfaction within the Terraform neighborhood?
The Kubernetes and Cloud Native Associate (KCNA, up 6.3%) was a shiny spot in IT certification. Whether or not Kubernetes is overly advanced (maybe as a result of it’s overly advanced) and whether or not or not firms are shifting out of the cloud, KCNA certification is a worthwhile asset. Cloud native purposes aren’t going away. And whether or not they’re managing Kubernetes complexity by constructing developer platforms, utilizing a Kubernetes supplier, or utilizing another resolution, firms will want folks on their employees who can show that they’ve Kubernetes expertise.
Cloud and cloud certifications
Content use for the most important cloud suppliers and their certifications was down throughout all classes, with one exception: Use of content material to organize for Google Cloud certifications is up 2.2%.
What does that inform us, if something? Are we taking a look at a “cloud repatriation” motion in full swing? Are our clients shifting their operations again from the cloud to on-prem (or hosted) information facilities? Last yr, we mentioned that we see little or no proof that repatriation is going on. This yr? An article in The New Stack argues that cloud repatriation is gathering steam. While which may account for the decline in the usage of cloud-related content material, we nonetheless see little proof that repatriation is definitely occurring. Two case research (37signals and GEICO) don’t make a development. The ongoing expense of working software program within the cloud most likely is larger than the price of operating it on-premises. But the cloud permits for scaling on demand, and that’s necessary. It’s true, few companies have the sudden utilization peaks which might be pushed by occasions like retail’s Black Friday. But the cloud suppliers aren’t nearly sudden 10x or 100x bursts of visitors; additionally they assist you to scale easily from 1x to 1.5x to 2x to 3x, and so forth. It saves you from arguing that you just want extra infrastructure till the necessity turns into a disaster, at which level, you don’t must develop 1.5x; you want 5x. After shifting operations to the cloud and experiencing just a few years of progress—even when that progress is reasonable—shifting again to an on-premises information middle would require important capital expense. It will most likely require gutting all of the infrastructure that you just haven’t been utilizing for the previous yr and changing it with one thing up-to-date.
Does this imply that cloud suppliers are “roach motels,” the place you may transfer in however you may’t transfer out? That’s not completely unfaithful. But the convenience of scaling by allocating just a few extra servers and seeing a barely increased invoice the subsequent month can’t be ignored, even when these barely increased payments sound just like the proverbial story of boiling the frog. Evaluating distributors, ready for supply, putting in {hardware}, configuring {hardware}, testing {hardware}—that’s effort and expense that companies are offloading to cloud distributors. The skill to scale fluidly is especially necessary within the age of AI. Few firms have the talents wanted to construct on-premises infrastructure for AI, with its cooling and energy necessities. That means both shopping for AI companies instantly from cloud suppliers or constructing infrastructure to host your personal fashions. And after all, the cloud suppliers have loads of assist for firms that want to make use of their high-end GPUs. (Seriously—if you wish to host your AI utility on-premises, see how lengthy it is going to take to get supply of NVIDIA’s newest GPU.) The actuality, as IDC concluded in a survey of cloud use, is that “workload repatriation from public cloud into dedicated environments goes hand in hand with workload migration to public cloud activities, reflecting organizations’ continuous reassessment of IT environments best suited for serving their workloads.” That is, there’s a relentless ebb and movement of workloads to and from public clouds as firms adapt their methods to the enterprise atmosphere.
The buzzword energy of “the cloud” lasted longer than anybody may moderately have anticipated, nevertheless it’s useless now. However, that’s simply the buzzword. Companies might not be “moving to the cloud”; that transfer has already occurred, and their employees not must learn to do it. Organizations now must learn to handle the investments they’ve made. They must study which workloads are most acceptable for the cloud and that are higher run on-premises. IT nonetheless wants employees with cloud expertise.
Security
Security Governance drove essentially the most content material use in 2024, rising 7.3% within the course of and overtaking Network Security (down 12%). The rise of governance is a vital signal: “Security” is not an advert hoc challenge, fixing vulnerabilities in particular person purposes or particular companies. That strategy results in limitless firefighting and finally failure—and people failures find yourself within the main information media and lead to executives dropping their jobs. Security is a company-wide challenge that must be addressed in each a part of the group. Confirming the rising significance of safety governance, curiosity in Governance, Risk, and Compliance (GRC) grew 44%, and Compliance grew 10%. Both are key elements of safety governance. Security structure additionally confirmed a small however important enhance (3.7%); designing a safety structure that works for a complete group is a vital a part of trying on the general safety image.
The use of content material about Application Security additionally grew considerably (17%). That’s a really common subject, and it maybe doesn’t say a lot besides that our customers are excited by securing their purposes—which matches with out saying. But what sorts of purposes? All of them: net purposes, cloud purposes, enterprise intelligence purposes, all the things. We get an even bigger sign from the rise in Zero Trust (13%), a very necessary technique for securing companies by which each person, human or in any other case, should authenticate itself to each service that it makes use of. In addition, customers should have acceptable privileges to do what they should do, and no extra. It’s notably necessary that zero belief extends authentication to nonhuman customers (different computer systems and different companies, whether or not inner or exterior). It’s a response to the “hard, crunchy outside, but soft chewy inside” safety that dominated the Nineteen Nineties and early 2000s. Zero belief assumes that attackers can get by way of firewalls, that they’ll guess passwords, and that they’ll compromise telephones and computer systems after they’re outdoors the firewall. Firewalls, good passwords, and multifactor authentication programs are all necessary—they’re the exhausting, crunchy outdoors that stops an attacker from getting in. Zero belief helps preserve attackers outdoors, after all—however greater than that, it limits the harm they’ll do as soon as they’re inside.
We’re puzzled by the drop in use of content material about Network Security, which corresponds roughly to the drop in Cisco certifications. Network Security remains to be the second most generally used talent, nevertheless it’s down 12% from 2023 to 2024. Perhaps community safety isn’t deemed as necessary when staff wander out and in of firm networks and purposes are distributed between in-house servers and the cloud. We hope that our customers aren’t making that mistake. An even bigger challenge is that networks haven’t modified a lot up to now few years: We’re nonetheless utilizing IPv4; we’re nonetheless utilizing routers, switches, and firewalls, none of which have modified considerably lately. What has modified is the way in which safety is carried out. Cloud computing and 0 belief have moved the main focus from big-iron networking gadgets to interactions between programs, no matter how they’re related.
Security certifications
Security certification has been one of many largest progress areas on our platform. As I’ve mentioned elsewhere, safety professionals love their certifications. There’s a great purpose for that. In most different specialties, it’s potential to construct a portfolio of applications you wrote, programs you architected, websites you’ve designed. What can a safety individual say in a job interview? “I stopped 10,000 people from logging in last year?” If you’ve ever monitored a public-facing Linux system, you understand that declare means little. Security is cursed with the issue that the very best information isn’t any information: “Nothing bad happened” doesn’t play effectively with administration or future employers. Neither does “I kept all the software patched, and spent time reading CVEs to learn about new vulnerabilities”—regardless that that’s a wonderful demonstration of competence. Certification is a manner of proving that you’ve sure expertise and that you just’ve met some widely known requirements.
The CISSP (up 11%) and CompTIA Security+ (up 13%) certifications are all the time on the high of our lists, and this yr isn’t any exception. Our State of Security in 2024 report confirmed that CISSP was the certification mostly required by employers. If there’s a gold customary for safety expertise, CISSP is it: It’s an intensive, complete examination for folks with greater than 5 years of expertise. CompTIA Security+ certification has all the time trailed CISSP barely in our surveys and in platform efficiency, however its place in second place is uncontested. Security+ is an entry-level certification; it’s notably fascinating for people who find themselves beginning their safety careers.
Security certification was particularly necessary for presidency customers. For most {industry} sectors, utilization targeted on programming expertise in Java or Python, adopted by synthetic intelligence. The authorities sector was a robust outlier. Security and IT certifications have been by far crucial matters. CompTIA Security+ and CISSP (in that order) led.
Moving past CISSP and Security+, most of the different safety certifications additionally confirmed features. Certified Ethical Hacker (CEH) was up 1.4%, as was the much less in style CompTIA PenTest+ certification (3.3%). Certified Cloud Security Professional was up 2.4%, considerably lower than we’d anticipate, given the significance of the cloud to fashionable IT, nevertheless it’s nonetheless a achieve. ISACA’s Certified in Risk and Information Systems Control (CRISC) was up 45%, Certified Information Security Manager (CISM) grew 9.3%, and Certified Information Security Auditor (CISA) was up 8.8%; these three certifications are strongly related to safety governance. The most important declines have been for the CompTIA Cybersecurity Analyst (CySA+) certification (down 13%) and CCNA Security (down 55%). The drop in CCNA Security is excessive, nevertheless it isn’t surprising on condition that not one of the Cisco certifications confirmed a rise this yr.
We’re lacking one necessary piece of the safety certification puzzle. There’s no information on AI safety certifications—and that’s as a result of there aren’t any. Software that comes with AI should be constructed and operated securely. That would require safety consultants with AI experience (and who can show that experience by way of certifications). We anticipate (or perhaps a greater phrase is “we hope”) that lack will likely be addressed within the coming yr.
Professional Development
Professional improvement continues to be an necessary progress space for our viewers. The most necessary talent, Professional Communication, grew 4.5%—not a lot however important. We noticed a 9.6% enhance in customers eager to know extra about Engineering Leadership, and a 21.5% enhance in customers utilizing content material about Personal Productivity.
Project Management was nearly unchanged from 2023 to 2024 (up 0.01%), whereas the usage of content material in regards to the Project Management Professional (PMP) certification grew 15%. Interest in Product Management declined 11%; it appears to be a talent that our customers are much less excited by. Why? For the previous few years, product supervisor has gave the impression to be a stylish new job title. But in final yr’s report, Product Management solely confirmed a small achieve from 2022 to 2023. Is curiosity in Product Management as a talent or as a job title fading?
We additionally noticed a 7.9% decline in Leadership (other than Engineering Leadership), and an enormous 35% decline for IT Management. Are we in charge these on the company layoff cycle? That’s potential, nevertheless it’s too straightforward. IT could also be affected by a common development towards simplification and platform engineering, as we’ve mentioned: A platform engineering group can do lots to cut back cognitive overhead for builders, nevertheless it additionally reduces the necessity for IT employees. A platform engineering group doesn’t should be massive; is the necessity for IT employees shrinking? The decline in Leadership could also be as a result of it’s a obscure, nonspecific time period, not like Engineering Leadership (which is up). Engineering Leadership is concrete and it’s one thing our engineering-oriented viewers understands.
New Initiatives
In 2024, we launched a number of new options on the O’Reilly studying platform, together with badges, quizzes, and a brand new model of O’Reilly Answers. What are they telling us?
Badges and Quizzes
We began a badging program late in 2023: Users from enterprise accounts can earn badges for taking programs and finishing quizzes. We gained’t go into this system particulars right here, however for the reason that program began, customers have earned practically 160,000 badges. We’re nonetheless constructing this system, however we’re inspired by its first yr.
Badges can provide us extra perception into what our customers are studying. The hottest badges are for Python expertise, adopted by GPT and immediate engineering. Generative AI and machine studying are additionally excessive on the listing. Kubernetes, regardless of its decline in models considered, was the fourth-most-frequently-acquired badge, with nearly the identical variety of badges earned as software program structure. Linux, SQL, skilled communication, and Java rounded out the highest 11. (Yes, 11—we needed to incorporate Java). The distinction between Java and Python is hanging, on condition that the usage of content material about these expertise is comparable. (Python leads Java, however not by a lot.) Oracle has a extremely regarded Java certification program, and there’s actually no equal for Python. Perhaps our customers acknowledge that getting a Java badge is superfluous, whereas acquiring badges for Pythonic expertise is significant?
Quizzes are intently tied to badges: If a ultimate quiz exists for a course or for a ebook, college students should cross it to earn their badge. Quiz utilization seems to comply with the identical traits as badging, although it’s untimely to attract any conclusions. While just a few legacy quizzes have been on the platform for a very long time (and aren’t related to badging), the push to develop quizzes as a part of the badging program solely started in June 2024, and quiz utilization remains to be as a lot a consequence of the time the quiz has been accessible on the platform as it’s of the talent for which it’s testing.
We may take a look at the experience required by the badges that have been earned. All of our content material is tagged with a talent degree: newbie, beginner-intermediate, intermediate, intermediate-advanced, or superior. 42% of the badges have been earned for content material judged to be intermediate. 33% of the badges have been earned for newbie content material, whereas solely 4.4% have been for superior content material. It’s considerably shocking that a lot of the badges have been earned for intermediate-level content material, although maybe that is sensible given the badge program’s B2B context: For essentially the most half, our customers are professionals moderately than freshmen.
Answers
One of our most necessary new options in 2024 was an improve to O’Reilly Answers. Answers is a generative AI-powered software that permits customers to enter pure language questions and generates responses from content material in our platform. Unlike most different generative AI merchandise, Answers all the time offers hyperlinks to the unique sources its responses are based mostly on. These citations are tracked and used to calculate writer royalties and funds to publishing companions.
So the plain query is: What are our customers asking? One would possibly guess that the questions in Answers can be much like the search phrases used on the platform. (At this level, Answers and search are distinct from one another.) That guess is partly proper—and partly improper. There are some apparent variations. Common search phrases embody ebook titles, writer names, and even ISBNs; titles and writer names hardly ever seem in Answers. The commonest searches are for single phrases, akin to “Python” or “Java.” (The common size of the highest 5,000 searches in September 2024 was two phrases, for example.) There are few single phrase questions in Answers (although there are some); most questions are well-formed sentences like “How many ways can you create a string object in Java?” (The common query size was 9 phrases.)
To analyze the questions from O’Reilly Answers, we basically turned them again into single-word questions. First, we eradicated questions from a “question bank” that we created to prime the pump, because it have been: Rather than requiring customers to jot down a brand new query, we provided a listing of prewritten queries they might click on on. While there’s undoubtedly some helpful sign in how the query financial institution was used, we have been extra excited by what customers requested of their very own volition. From the user-written questions, we created an enormous “bag of words,” sorted them by frequency, and eradicated stopwords. We included a number of stopwords that aren’t in most lists: phrases like “data” (what does that imply by itself?) and “chapter” (sure, you may ask a couple of chapter in a ebook, however that doesn’t inform us a lot).
With that background in thoughts, what have been the most typical phrases in Answers and in searches? In order:
| Answers | Search Queries |
| Python | Python |
| Java | Machine studying |
| Management | Kubernetes |
| Key | Java |
| Model | Rust |
| Security | React |
| File | AWS |
| Architecture | CISSP |
| AI | C++ |
| System | Linux |
| Service | Docker |
| Project | SQL |
| Learning | JavaScript |
There’s an apparent distinction between these two lists. The Answers listing consists largely of phrases that could possibly be a part of longer questions. The Search listing is made up of matters and expertise about which one would possibly need data. That’s hardly shocking or insightful. We’ve mentioned most searches on the platform are single-word searches, which implies that these phrases should be stand-alone expertise or matters, like Python or Java. Likewise, Answers was constructed to permit customers to ask extra detailed, in-depth questions and get targeted solutions from the content material on our platform—so moderately than seeing single phrase searches, we’re seeing frequent phrases from longer questions. Maybe that’s a self-fulfilling prophecy, nevertheless it’s additionally displaying that Answers is working the way in which we supposed.
There’s a little bit extra sign right here. Python and Java are the 2 high programming languages on each lists, but when we take a look at search queries, machine studying and Kubernetes are sandwiched between the 2 languages. That could be a results of our customers’ experiences with companies like ChatGPT. Programmers rapidly realized that they’ll get cheap solutions to questions on Java and Python, and the prompts don’t should be very advanced. My private favourite is “How do you flatten a list of lists in Python?,” which may be answered by most chatbots accurately however isn’t significant to our search engine.
Kubernetes raises a special query: Why is it the third-most-common search engine question however doesn’t seem among the many high phrases on Answers? (It’s the Ninetieth-most-common phrase on Answers, although the precise rank isn’t significant.) While Kubernetes is a subject that’s amenable to express questions, it’s a fancy software, and developing with exact prompts is tough; writing a great query most likely requires a great understanding of your IT infrastructure. You would possibly want to grasp remedy your drawback earlier than you may ask a great query about remedy your drawback. A search engine doesn’t face issues like this. It doesn’t want extra data to return a listing of sources.
Then what about phrases like Rust and Linux, that are excessive on the listing of frequent searches, however not within the high 13 for Answers? It’s comparatively straightforward to give you particular questions on both of those—or, for that matter, about SQL, AWS, or React. SQL, AWS, and Linux are moderately near the highest of the Answers glossary. If we simply concern ourselves with the order by which phrases seem, issues begin to fall into place: AWS (and cloud) comply with studying; they’re adopted by Linux, adopted by SQL. We’re not stunned that there are few questions on CISSP on Answers; it’s a certification examination, so customers usually tend to need check prep materials than to ask particular questions. Rust and React are nonetheless outliers, although; it’s straightforward to ask exact and particular questions on both of them. Rust remains to be unfamiliar to lots of our customers—may the reason be that our clients wish to study Rust as a complete moderately than ask particular questions which may solely happen to somebody who’s already realized the language? But in case you settle for that, React nonetheless stays an outlier. We might know the solutions subsequent yr, at which period we’ll have a for much longer observe report with Answers.
The Coming Year
That wraps up final yr. What will we see this yr? We’ve given hints all through this report. Let’s deliver all of it collectively.
AI dominated the information for 2024. It will proceed to take action in 2025, regardless of some disillusionment. For essentially the most half, those that are disillusioned aren’t the folks making selections about what merchandise to construct. While concern about jobs is comprehensible in a yr that’s seen important layoffs, we don’t imagine that AI is “coming for your job.” However, we do imagine that the long run will belong to those that learn to use AI successfully—and that AI could have a profound impression on each occupation, not simply IT and never simply “knowledge workers.” Using AI successfully isn’t nearly developing with intelligent prompts so you may copy and paste a solution. If all you are able to do is immediate, copy, and paste, you’re about to grow to be superfluous. You want to determine work with AI to create one thing that’s higher than what the AI may do by itself. Training staff to make use of AI successfully is without doubt one of the greatest issues an organization can do to organize for an AI-driven future. Companies that don’t spend money on coaching will inevitably fall behind.
In the approaching yr, will firms construct AI purposes on high of the enormous basis fashions like GPT-4, Claude, and Gemini? Or will they construct on high of smaller open fashions, lots of that are based mostly on Meta’s Llama? And within the latter case, will they run the fashions on-premises (which incorporates the usage of internet hosting and colocation suppliers), or will they lease use of those open AI fashions as a service from varied suppliers? In the approaching yr, watch fastidiously what occurs with the small open fashions. They already ship efficiency nearly pretty much as good as the inspiration fashions and can undoubtedly be the premise for a lot of AI purposes. And we suspect that almost all firms will run these fashions within the cloud.
Security is the opposite important progress space. Companies are waking as much as the necessity to safe their information earlier than their reputations—and their backside strains—are compromised. Waking up has been an extended, sluggish course of that has sunk the careers of many CEOs and CIOs, nevertheless it’s occurring. Our customers are learning to realize safety certifications. We see firms investing in governance and placing in company-wide insurance policies to take care of safety. In this respect, AI cuts each methods. It’s each a software and a hazard. It’s a software as a result of safety professionals want to observe over big streams of knowledge, on the lookout for the anomalies that sign an assault; it’s a software as a result of AI can digest sources of details about new threats and vulnerabilities; it’s a software as a result of AI can automate routine duties like report technology. But it’s additionally a hazard. AI-enabled purposes enhance a corporation’s risk floor by introducing new vulnerabilities, like immediate injection, that we’re solely now studying mitigate. We haven’t but seen a high-profile assault towards AI that compromised a corporation’s skill to do enterprise, however that can definitely occur finally—perhaps in 2025.
Whatever occurs this yr, AI will likely be on the middle. Everyone might want to learn to use AI successfully. AI will inevitably reshape all of our professions, however we don’t but understand how; we’re solely beginning to get glimpses. Is that thrilling or terrifying? Both.
Footnotes
- The definition of “open” and “open source” for AI remains to be controversial. Some open fashions don’t embody entry to weights, and lots of don’t embody entry to coaching information.