50,000 People at the ‘Super Bowl of Software’ Told CNBC on 18 September That Last Year’s AI Is Enough — the 3 AI Engineer Job Descriptions I Rewrote in Singapore This Week, and the Frontier Skills I Stopped Paying For

The most useful thing published about AI this week was not a keynote. It was a report filed by CNBC on the morning of Thursday 18 September from the exhibition floor of Dreamforce, Salesforce’s 50,000-person conference in San Francisco, and its headline was almost boring: “AI safety debate meets reality at Dreamforce as business leaders say last year’s models are enough.” I read it on the MRT on Friday morning with three open AI engineer requisitions for Singapore clients in my inbox, all three asking for some version of “experience at the frontier”. By Friday evening I had rewritten all three. This is the report, what it says about how AI is actually being used by the companies that pay for it, and the hiring change I think every Singapore employer building on top of models should make this month.

What the Floor Said While the Stage Argued

The stage at Dreamforce this year was a debate about speed. On Tuesday 15 September, Anthropic’s Dario Amodei, following an essay urging the industry to slow model development, shared the keynote with Salesforce CEO Marc Benioff; Nvidia’s Jensen Huang, on the same stage, urged frontier labs to “run as fast as you can”. OpenAI’s Sam Altman appeared too. The backdrop, as CNBC noted, was an Anthropic researcher who had resigned days earlier saying the top labs were “gambling with our lives”, and safety proposals from both Amodei and Altman that followed.

The floor was somewhere else entirely. CNBC’s reporters went to the people who buy and deploy the software, and got a consistent answer that had nothing to do with either speed or existential risk. The quotes are worth reading in full, because each one is a hiring decision in disguise.

• Alec Bronston, senior Salesforce director at the Chicago retail-data company Spins: “It’s already hard enough to keep up.” A slowdown would offer “a lot of opportunity to just even catch up and get our feet wet”.
• Jaya Rohit Vuyyuru, vice president at the consulting firm SummitX: “The frontier models are way ahead already. A lot of the customer base is still getting their feet wet. There’s still that gap where the clients can still fulfil and start getting a sense of what agents can do.”
• Tim Sanders, chief innovation officer at G2: “The majority of agentic outcomes aren’t driven by frontier capabilities. They’re driven by last year’s AI. It’s not that relevant to agentic providers, and certainly not that relevant to SaaS.”
• Kevin Lee, technology chief at the contact-centre software vendor Nice: “In large part, with the models that are out there already today, and even one generation behind, they are highly performant and effective at doing the things that our customers need. It’s almost like everything beyond this point is icing on the cake.”
• Allan Thygesen, CEO of Docusign: the company uses “all the big frontier models as well as some of the open-weight models” and routes each request to the most cost-effective system. Docusign’s own engineering blog says larger frontier models are “reserved for judgment-intensive work — complex clause analysis, multi-document reasoning, summarization — where general reasoning capability is worth the higher per-call cost”.

Two more details from the piece matter. Salesforce’s own Agentforce bots, according to a support page CNBC cited, are not built on Anthropic’s newest Claude Fable 5.1 or OpenAI’s new GPT-6 Astra. And Sanders’ second point, about the token economy, is the one Singapore finance directors will care about: “SaaS up to now has had no variable cost to deliver services.” Now it does, and every AI feature a company ships has a cost per call that somebody has to engineer down. We covered Salesforce’s product side of this event in our note on the seven job-ready Agentforce agents; this piece is about what the buyers said, which is more useful.

31 Singapore AI Requisitions, 19 Asked for the Frontier, 3 Needed It

Before rewriting anything I went back through the AI engineer roles we have filled in Singapore since January. Thirty-one requisitions across fintech, logistics, healthtech, two government-linked teams and a scattering of SaaS companies. I coded each one two ways: what the job description asked for, and what the engineer, six months in, actually spends their week doing. The gap is the chart below, and it is the gap CNBC described.

Nineteen of the thirty-one asked for some variant of frontier-model research: “experience training or fine-tuning large language models”, “deep familiarity with the latest frontier releases”, “publications or open-source contributions in LLM research”. Three of those nineteen roles, all in teams that fine-tune open-weight models on proprietary data, use that experience. The other sixteen engineers spend their time on exactly the work the Dreamforce buyers described: choosing which model handles which request, building the evaluations that make that choice defensible, watching token cost per outcome, wiring the model into the CRM, the ticketing system and the data warehouse the company already runs, and writing guardrails. That is applied engineering, and the requisitions were priced for research.

The 3 Job Descriptions I Rewrote This Week

All three were live requisitions for Singapore clients on Friday morning. None of the clients trains models. I show the before and after because the change is mostly deletion, and deletion is what employers find hardest.

1. “Senior LLM Research Engineer” became “AI Platform Engineer (Routing & Evaluation)”

The client is a logistics SaaS company with a customer-support assistant and a document-extraction pipeline in production. The original JD asked for “experience training or fine-tuning LLMs at scale”, “strong publication record or open-source contributions” and “deep familiarity with frontier model architectures”. The engineer they actually have, who is leaving, spent this year building a router that sends 80-odd percent of support requests to a small model, escalates ambiguous ones to a mid-tier model, and reserves the frontier model for multi-document disputes, exactly the Docusign pattern. The new JD asks for: owning the routing layer and its evaluation set; a demonstrated record of reducing cost per resolved ticket while holding a quality metric; experience deploying at least one open-weight model behind an API; and the ability to explain a routing decision to a product manager. Salary band moved from the research tier to the senior applied tier, roughly 20 percent lower, and the shortlist went from four candidates to seventeen in three days.

2. “Frontier Model Specialist” became “Applied AI Engineer, Evaluation Owner”

A healthtech client with a clinical-notes summariser. The original JD wanted somebody who “tracks and benchmarks every frontier release within a week of launch”. What the team needs, and what Kevin Lee’s line about “one generation behind” makes clear, is not somebody who chases releases but somebody who can say, with evidence, whether a new release is worth the migration cost for this workload. The new JD centres on the evaluation harness: building the golden set with clinicians, running every candidate model against it, reporting quality, latency and cost per summary side by side, and making the upgrade-or-not call quarterly. The frontier-tracking line survived as a nice-to-have. Our guide to hiring LLM evaluation engineers in Singapore has the interview loop for this profile; it is the one we now run.

3. “AI Solutions Architect” became “AI Solutions Architect (Cost-to-Serve)”

A fintech client building agent workflows into an existing Salesforce and Slack estate. The title stayed, the requirements changed. Gone: “hands-on experience with the latest agentic frameworks from the major labs”. Added: a cost-to-serve model for every workflow before it ships, token budgets per agent with alerts, and the integration work that Sanders was pointing at, meaning the agent lives where the staff already work rather than in a new tab. This is the role where Sanders’ point about variable cost bites hardest: a finance team that has never had a per-call cost on a software feature now needs somebody who can forecast it. Our cost-effective AI engineering team guide has the budgeting method; the Salesforce developer hiring guide covers the integration side.

The 4 Frontier Skills I Stopped Paying For, and the 5 I Now Pay For

To be precise about the deletion. These are the requirement lines I now strike from any Singapore requisition where the company is building on hosted or open-weight models rather than training its own.

• “Experience pre-training or large-scale fine-tuning”. Relevant to perhaps a dozen teams in Singapore. If you are one of them, keep it. If you are not, it filters out most of the people who can do the job and adds S$3,000 to S$5,000 a month to the band.
• “Publications in top ML venues”. A signal of research ability, which is not the job.
• “Deep familiarity with the latest frontier releases”. Everyone competent reads the release notes. What you want is somebody who can measure a release against your golden set, which is a different requirement.
• “Expert in [specific lab]’s agent framework”. The framework will change. The routing, evaluation and integration skills transfer; the framework knowledge is a week’s reading.

And the five lines that go in, in the order they appear in the rewritten JDs.

1. Model routing. “Has designed and operated a tiered routing layer across at least two model providers or an open-weight deployment, with measured cost and quality outcomes.”
2. Evaluation ownership. “Has built and maintained a golden evaluation set with domain experts and used it to make an upgrade, downgrade or vendor-switch decision.”
3. Token-cost engineering. “Can produce a cost-to-serve per outcome for an AI feature before launch and instrument it after.”
4. Open-weight deployment. “Has deployed at least one open-weight model behind a production API, including the serving, scaling and monitoring.” Our open-weight deployment evaluation guide is the interview for this line.
5. Integration. “Has shipped an AI feature inside an existing system of record (CRM, ticketing, ERP, chat) rather than as a standalone app.”

What It Does to the Salary Band in Singapore

The honest number from our placements this year: applied AI engineers with the five skills above have sat roughly 15 to 25 percent below candidates with genuine frontier-research experience, at the same seniority. The research premium is real and, for the teams that need it, worth paying. In sixteen of our thirty-one requisitions it was being paid for skills the role does not use, which is the single most expensive line item in Singapore AI hiring that nobody talks about. The rewritten profile also fixes the pipeline problem: the frontier-research pool in Singapore is small and heavily contested by the labs, the sovereign-AI programmes and the banks, while the applied pool is roughly four times larger and includes strong backend engineers who have crossed over in the last eighteen months. Our AI engineer salary guide for Singapore has the bands by seniority; the salary calculator lets you run the two versions of a role side by side.

If You Also Hire AI Engineers in Dubai

The same rewrite applies, with one difference: the UAE’s sovereign-AI programmes create more genuine demand for research profiles than Singapore does, so the frontier line belongs in more requisitions there. Our Dubai team’s guide to writing AI engineer job descriptions in Dubai has the structure, and their note on hiring multi-model AI engineers in Dubai is the routing profile from the other side of the Indian Ocean, written for a market where vendor diversification is a policy question as much as a cost one.

FAQ — Dreamforce, Last Year’s Models and AI Hiring in Singapore