[AINews] Anthropic ends Coding Interviews: "Hunger beats tenure"

Felix Rieseberg, a hiring manager at Anthropic and a past guest of the publication, said this month that he has retired the traditional coding interview for the candidates he evaluates. His reasoning: Claude is now better than people at both a small React exercise and assembly-level quizzing, so neither test tells him much. The builders who make the largest impact, he said, succeed regardless of how well they perform on such screens. He summarized the shift in a phrase: hunger beats tenure.
The announcement arrived in a roundup heavy with incremental releases. Google's Genie world model, Artificial Analysis's model index and Cognition's FrontierCode benchmark each shipped a major version bump, alongside work on formal math verification. Rieseberg's post carried the headline because of its implications for engineers entering the field rather than for any new capability.
His advice for juniors rests on a handful of principles. The first is to optimize for business impact rather than the ticket backlog, citing Patrick McKenzie's 2011 line that engineers are hired to create business value, not to program things. The second is to pick a radical idea, one that if it worked would dramatically improve the organization's chances of hitting its goals. The third is to build it over a weekend: he has watched small groups and solo engineers skip the usual product requirements document, the buy-in rounds and the lobbying for designer or product manager time simply by arriving with a working prototype.
The fourth is discipline in scope. In his experience juniors often stuff several ideas into one prototype, burying three excellent ones among twenty bad ones; better to make one idea legible. The fifth is to exploit a fresh perspective. On user-facing work, newer engineers often carry more empathy for users than anyone else on the team and can sense what works and what does not.
This reads as a hiring philosophy shaped by cheap code generation. When producing a working feature is fast, the scarce inputs become judgment about what to build and evidence that it should exist. A prototype is also a better signal than a whiteboard exercise because it demonstrates taste, sequencing and follow-through at once, which interviews rarely capture. The obvious trade-off is that weekend prototypes can reward showmanship over maintainable engineering, and a candidate who is strong at systems or debugging but weak at product framing may be harder to evaluate under this model. Rieseberg's post is a managerial position rather than a company-wide policy statement, and it leaves open how Anthropic assesses depth in areas a quick prototype cannot reveal.
The same roundup carried heavy technical news elsewhere. Google DeepMind introduced Genie 4, a real-time world model built on Gemini Omni with Gemini planning multi-step events in the loop. It adds native ambient audio, music and character dialogue, a first for the Genie line, and DeepMind claims improvements in aerodynamics, collisions and natural human motion, responding to prompted events mid-generation rather than producing static worlds. The research preview is open to trusted testers now, with Google AI Ultra subscribers next through Project Genie, then early-access partners via the Gemini API and AI Studio, marking Genie's first API availability.
On the research side, Axiom announced a complete Lean auto-formalization of the Classification of Finite Simple Groups, a proof spanning roughly 15,000 pages across about 500 papers by more than 100 mathematicians. An agent swarm produced about 50 million lines of Lean in 24 days, with agents claiming tasks, reviewing each other and merging through GitHub under human guidance; the effort reportedly fixed hundreds of errata. OpenAI announced results including a claimed refutation of the Exponential Time Hypothesis, and Anthropic reportedly says Claude independently disproved it, finding the key algorithm on Oct 16 with 14 million output tokens and no human input. Both are claims awaiting independent scrutiny.
Artificial Analysis shipped Intelligence Index v5, with 60 percent of the index drawn from private or held-out sets. It adds FrontierCode 2.0 and Terminal-Bench-Science, swaps HLE for HLE-Diamond, drops GDPval-AA, SciCode and AA-LCR, and weights four categories at 25 percent each. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Gemini 4 Argon at 50. The open-weights gap widened from 3 to 12 points. Cognition's FrontierCode 2.0 uses deliberately underspecified prompts and grades whether a maintainer should merge the change; the top score is 45 percent, and about 80 percent of submitted pull requests were rejected during construction.
Microsoft's MAI-Image-2.7-Preview ranks second on AA-Image-Editing behind GPT Image 2.5 Sunburst and first on four of seven editing actions, available for editing only in MAI Playground. Ling 3.1 Flash 1019 is an open model with 560 billion total and 25 billion active parameters, a 1M context and an MIT license, scoring 44 on AA Index v4.3, up from 41, priced at $0.30/$0.90 per million input/output tokens. Liquid released six roughly 2B open-weight models on a tri-hybrid architecture optimized across 14 processor classes and fine-tunable on one GPU, including LFM3-1.7B with 256K context and vision, audio and omni variants.
Why it matters: Rieseberg's stance signals that at the frontier labs, the coding screen is losing value as a filter just as AI assistants get good at exactly those exercises, which shifts what junior candidates must demonstrate. For developers entering the field, the practical change is that a shipped, narrowly scoped prototype with a clear business rationale may carry more weight than algorithm drills, an inference from his advice rather than a stated hiring rule. Product and engineering leaders reading the same signal may reconsider their own interview loops, though the model assumes candidates can build quickly with capable tools and says nothing about how to evaluate deep systems skill. The surrounding releases, from Genie 4's API access to the wider open-weights gap on the new index, suggest the tooling that makes weekend prototypes feasible is still improving fast.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Since coding is solved, why code for the job?