This article is crossposted from IEEE Spectrum’s careers newsletter. Sign up now to get insider tips, expert advice, and practical strategies, written in partnership with tech career development company Parsity and delivered to your inbox for free!

There is no shortage of people telling recent engineering graduates that their degree was a mistake and that AI is coming for their jobs before they even land one. I respectfully disagree.

I have been a software engineer for 12 years, done well over 100 interviews on both sides of the table, and run Parsity, an AI engineering program. A few patterns emerge consistently in who actually breaks through in today’s job market. Here’s why I think the job market isn’t as dire as it looks, and what I would do if I were looking for my first tech job.

The Numbers Need Context

The Federal Reserve Bank of New York recently placed unemployment for recent CS graduates in the United States at 6.1 percent, with computer engineering graduates at 7.5 percent. Compared to philosophy majors at 3.2 percent and art history graduates at 3.0 percent, those figures look alarming. They require more context than most headlines provide.

When researchers factor in underemployment (graduates working jobs that don’t require a college degree), then engineers are doing relatively well, coming in below 20 percent, against a 42 percent average across all recent graduates. Many majors reporting lower unemployment are achieving that figure by accepting work entirely unrelated to their field. Scored across unemployment, underemployment, and early-career earnings together, CS and computer engineering still rank among the top fields for overall labor market outcomes.

The degree is not the problem. The hiring pipeline is. Job postings labeled “entry-level software engineer” grew roughly 47 percent between late 2023 and late 2024, while actual hiring into those roles dropped approximately 73 percent in the same window. So-called “ghost jobs,” used to create an illusion of company growth, are everywhere. This makes the front door harder to find, but it exists.

Here Is What To Do About It

Do a broad search of your (real-life) network. Roughly 26 percent of job offers come through referrals. Look at your actual network—classmates, professors, past internship contacts, relatives—and identify people at companies that might be hiring. The goal is a warm introduction to someone who is or knows a decision maker. One introduction carries more weight than a hundred cold applications through a portal.

Find symmetric risk. A junior engineer is a risky hire by definition. A startup carries a matching risk profile, meaning potentially lower compensation, no certainty of longevity, and higher performance expectations. But that shared risk creates mutual interest. The learning curve is steep, the exposure is broad, and the track record transfers directly. For engineers whose longer-term goal is a large organization, a startup is not a detour. It can be how you build the experience those organizations eventually want to see. The first job is for validation and learning. It is not a life sentence.

Manufacture experience rather than waiting for it. Employers want experience but will not hire you to get it. The way through is to create it: a deployed project, an open-source contribution, building something real for a small business or family member. Recruiters are skeptical of toy projects. A deployed application solving a real problem, combined with the ability to talk clearly about the decisions you made and why, still moves the needle.

Gain practical AI engineering skills, not just AI tool fluency. Using Cursor or Copilot is now a baseline expectation. What differentiates candidates is going one level deeper. Most working engineers, including senior ones, have not built a RAG pipeline or designed a multi-agent system. Understanding how to chunk documents, generate embeddings, store and query them from a vector database, and wire it into a production application puts a candidate ahead of a significant portion of the market on a skill in rapidly growing demand. AI and data science roles grew 163 percent in job postings in 2025. The engineers who understand how these systems actually work, not just how to prompt them, are in the shortest supply.

Stop optimizing around conditions you cannot predict. Nobody anticipated the 2021 hiring boom. Nobody predicted this correction. Build durable skills. The demand for engineers who can reason clearly about systems is not going away. Where you start is not where you end.

—Brian

More major workforce reductions are on the horizon at Big Tech companies: Meta announced it will cut 10 percent of its workforce, or about 8,000 employees, and Microsoft plans to offer buyouts for 7 percent of its U.S. employees in a voluntary retirement program. The cuts are understood by many to be linked to AI. But is AI really to blame? For The Conversation, two academics at the University of Sydney give their two cents.

Read more here.

Tom Burick got his start as a roboticist. But when a financial downturn forced him to close his robotics business, he thought of the effect teachers had on his life and decided to pay it forward. Burick now works as a technology instructor at a school for students with autism, where he recently led a project building a full-scale replica of ENIAC, an historic computer celebrating its 80th anniversary this year.

Read more here.

Across several industries, the United States has been moving toward limiting the use of sensitive technology made in China. Now, legislation has been introduced to extend the trend to ground robots, including humanoids, dogs, and crawlers. This could benefit some U.S.-based robotics firms—but many of these companies still rely on Chinese-made components. “The U.S. robotics industry is in a pickle,” writes Spectrum tech policy editor Lucas Laursen.

Read more here.

Legora is already hiring to fill its new offices, with plans to grow its combined EMEA headcount to 700.

Swedish legal AI start-up Legora is expanding its footprint with a new engineering hub in London alongside new offices in Madrid, Milan and Paris. The expansion comes just months after the company raised $550m in a Series D round which valued Legora at $5.5bn.

Founded as Leya in 2023, Legora is an agentic AI platform supporting legal professionals with research, review and document drafting. It is used by more than 100,000 legal professionals at more than 1,200 law firms and in-house legal teams across more than 50 markets, according to the company.

The Madrid, Milan, and Paris offices will serve as regional hubs for customer success, go to market functions and legal engineering, while the new London engineering hub will be co-located with Legora’s existing presence in the city.

The new offices will open during Q3 this year and represent Legora’s most concentrated Europe, Middle East and Africa (EMEA) investment to date, the company said.

Legora is already hiring for positions to fill the new offices, with plans to grow its combined EMEA headcount to 700 within the next year. According to its website, Legora currently employs more than 400.

“Our customers in these countries have built Legora into the way they work,” said Max Junestrand, the CEO and co-founder of Legora. “Opening offices in Madrid, Milan and Paris means we can be genuinely close to them as we build the future of the platform together.

“Engineers who understand how AI applies in professional contexts are disproportionately concentrated in London,” said Junestrand added. “People here have built things that have to perform under real legal and regulatory constraints. That’s a different problem from building a consumer product, and it’s precisely the problem we’re solving.”

The new offices will bring Legora’s global footprint to 16 cities including Munich, Chicago, Houston, San Francisco, Toronto, Bengaluru and Sydney, as well as the recently announced offices in Singapore and Tokyo, the company said. Legora also has existing engineering hubs in Stockholm and New York.

Legora’s March Series D round was led by Accel, with participation from the likes of Benchmark, general Catalyst, Y Combinator, Menlo Ventures and Salesforce Ventures – taking the legal-tech’s total raise to date to $815m.

The company, at the time, said it planned to use the newly raised funds to further expand across the US, including with new offices in Texas and Illinois, as well as new local hubs. Legora plans to expand its US headcount to more than 300 by the end of 2026.

In April, the company acquired Stockholm-based AI-native legal research start-up Qura for an undisclosed value.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Engineering teams building agentic coding pipelines now have a concrete open-source alternative to managed models like Claude Fable 5 — one that runs on a single H100. The tradeoff: Cohere’s North Mini Code, which launched Tuesday, generated three times the output tokens of comparable models in independent testing, a verbosity cost that compounds in high-volume production workloads.

The new open-source model is a 30 billion parameter mixture-of-experts (MoE) model with 3 billion parameters active per token, built for agentic software engineering including sub-agent orchestration, architecture mapping, code review and terminal work. The model supports a 256,000 token context window with a 64,000 token maximum generation length, and is available on Hugging Face under an Apache 2.0 license.

What North Mini Code can do

North Mini Code targets the full agentic coding stack. Here is what the model does and what it runs on.

Software engineering. Cohere built North Mini Code specifically for agentic software engineering, not adapted from a general-purpose base. It has integrated tool-use capabilities and supports interleaved thinking, which Cohere says improves performance across multi-step agentic work.

Architecture mapping and code review. North Mini Code can analyze and map systems architecture, surface dependencies and perform code review across large codebases. With a 256,000 token context window, it can hold substantial multi-file projects in a single context pass.

Terminal-based agentic tasks. The model is trained for terminal environments, handling shell interactions, package scripts and command-line tooling. Cohere benchmarked it on Terminal-Bench v2, which tests agents in real terminal environments rather than synthetic code generation tasks.

How it was built

North Mini Code is a sparse mixture-of-experts model with 128 experts, of which 8 activate per token. The compute requirement at inference time is closer to a 3 billion parameter model despite 30 billion total parameters. Nick Frosst, co-founder of Cohere, demoed it running on a Mac Studio via MLX at around 20 gigabytes of RAM, the same machine he uses for his own local coding work.

Cohere trained the model through two stages of supervised fine-tuning followed by reinforcement learning with verifiable rewards across more than 70,000 verifiable tasks spanning approximately 5,000 repositories, deduplicated against SWE-Bench. 

Rather than optimizing against a single agent scaffold, Cohere trained across three. SWE-Agent uses a rich CLI with specialized commands. Mini-SWE-Agent uses a single bash tool with raw shell output. OpenCode uses individually typed tools returning structured JSON. Cohere reports a 10 percentage point gain on OpenCode evaluation from the multi-harness approach while maintaining SWE-Agent performance.

Where it fits

North Mini Code enters a market that now includes Mistral Devstral Small 2, GitHub Copilot, Cursor, and Claude Fable 5 — each with distinct cost and deployment tradeoffs.

Cohere’s primary benchmark comparison is against Mistral Devstral Small 2, a 24 billion parameter dense model. In vendor-reported internal tests, Cohere claims 2.8x higher output throughput and a 30% inter-token latency advantage over Devstral Small 2 in internal tests under identical hardware configurations. Cohere also claims, in its Hugging Face technical post, that North Mini Code outperforms open-source models up to four times its parameter count on its reported benchmarks, including models at 120 billion parameters.

Artificial Analysis independently ranks it eighth of 127 comparable open-weight models on output speed at 210 tokens per second, with a time to first token of 0.25 second against a class median of 1.95 seconds. It places 18th of 127 on the Artificial Analysis Intelligence Index. One flag from the same data: the model generated 75 million output tokens to complete the Intelligence Index against a class median of 25 million. In high-volume agentic pipelines, that verbosity compounds into inference cost and latency.

“Suddenly people are thinking like hey, am I getting enough economic value out of the tokens from a model?” Frosst said during the launch video. “Local deployment is one way of empowering people and making AI really something that works for them.”

GitHub Copilot, Cursor and Claude Code operate on per-usage or subscription pricing with no on-premises option. Anthropic’s Claude Fable 5, now the most capable publicly available managed coding model, runs at $50 per million output tokens. For Frosst, the model is the polar opposite of Fable.

“Its small, cost effective, apache 2.0, and locally deployable. This is the way LLMs should go. small, open source, transparent and sovereign, vs large, expensive, proprietary and hegemonic,” Frosst wrote in a post on X.

What this means for enterprises

For teams building production agentic coding pipelines, North Mini Code’s release clarifies a set of decisions that have been forming for months.

Purpose-built agentic training is now a baseline to evaluate against. The distinction between models fine-tuned for code and models trained specifically for agentic workflows, with verified tool calls and multi-harness robustness, is now a material factor in pipeline decisions. Any model vendor claiming agentic coding capability should be able to answer whether its training used verifiable agentic tasks or was adapted from a general-purpose base.

Verbosity is a hidden pipeline cost that benchmarks do not surface. Artificial Analysis measured North Mini Code generating three times the output tokens of comparable models. That verbosity compounds across inference cost and latency in high-volume pipelines. Throughput testing against actual workload volume is the evaluation step the benchmark rankings skip.

The frontier pricing split is now a real architectural decision. Fable 5 at $50 per million output tokens and North Mini Code on a single H100 represent a genuine tradeoff between cost control and data residency on one side, and managed infrastructure overhead on the other. Teams running high-volume agentic coding pipelines should model both cost paths against their actual workload before committing to either.

AI is rapidly expanding across finance, but most agentic offerings have yet to reach core production systems. Only 10% of enterprises are using AI tools in a meaningful, production-grade way. Not because of a lack of interest, but because connecting AI to core systems to trade capture, risk, and surveillance is still a work in progress.

These systems offer the greatest opportunity for AI to simplify finance operations through efficient workflows and live trading queries. Yet, legacy systems force this technology to operate in isolation. The volume of architecture connected to traditional platforms often creates this constraint.