Be part of the occasion trusted by enterprise leaders for almost 20 years. VB Rework brings collectively the individuals constructing actual enterprise AI technique. Be taught extra
Apple’s machine-learning group set off a rhetorical firestorm earlier this month with its launch of “The Phantasm of Pondering,” a 53-page analysis paper arguing that so-called giant reasoning fashions (LRMs) or reasoning giant language fashions (reasoning LLMs) equivalent to OpenAI’s “o” sequence and Google’s Gemini-2.5 Professional and Flash Pondering don’t truly have interaction in unbiased “considering” or “reasoning” from generalized first ideas realized from their coaching knowledge.
As a substitute, the authors contend, these reasoning LLMs are literally performing a form of “sample matching” and their obvious reasoning capacity appears to collapse as soon as a job turns into too advanced, suggesting that their structure and efficiency isn’t a viable path to bettering generative AI to the purpose that it’s synthetic generalized intelligence (AGI), which OpenAI defines as a mannequin that outperforms people at most economically priceless work, or superintelligence, AI even smarter than human beings can comprehend.
ACT NOW: Come talk about the most recent LLM advances and analysis at VB Rework on June 24-25 in SF — restricted tickets accessible. REGISTER NOW
Unsurprisingly, the paper instantly circulated broadly among the many machine studying group on X and plenty of readers’ preliminary reactions had been to declare that Apple had successfully disproven a lot of the hype round this class of AI: “Apple simply proved AI ‘reasoning’ fashions like Claude, DeepSeek-R1, and o3-mini don’t truly motive in any respect,” declared Ruben Hassid, creator of EasyGen, an LLM-driven LinkedIn publish auto writing software. “They only memorize patterns rather well.”
However now right now, a brand new paper has emerged, the cheekily titled “The Phantasm of The Phantasm of Pondering” — importantly, co-authored by a reasoning LLM itself, Claude Opus 4 and Alex Lawsen, a human being and unbiased AI researcher and technical author — that features many criticisms from the bigger ML group concerning the paper and successfully argues that the methodologies and experimental designs the Apple Analysis crew used of their preliminary work are basically flawed.
Whereas we right here at VentureBeat usually are not ML researchers ourselves and never ready to say the Apple Researchers are mistaken, the talk has definitely been a full of life one and the difficulty concerning the capabilities of LRMs or reasoner LLMs in comparison with human considering appears removed from settled.
How the Apple Analysis examine was designed — and what it discovered
Utilizing 4 basic planning issues — Tower of Hanoi, Blocks World, River Crossing and Checkers Leaping — Apple’s researchers designed a battery of duties that pressured reasoning fashions to plan a number of strikes forward and generate full options.
These video games had been chosen for his or her lengthy historical past in cognitive science and AI analysis and their capacity to scale in complexity as extra steps or constraints are added. Every puzzle required the fashions to not simply produce an accurate remaining reply, however to elucidate their considering alongside the way in which utilizing chain-of-thought prompting.
Because the puzzles elevated in problem, the researchers noticed a constant drop in accuracy throughout a number of main reasoning fashions. In probably the most advanced duties, efficiency plunged to zero. Notably, the size of the fashions’ inside reasoning traces—measured by the variety of tokens spent considering by way of the issue—additionally started to shrink. Apple’s researchers interpreted this as an indication that the fashions had been abandoning problem-solving altogether as soon as the duties turned too onerous, primarily “giving up.”
The timing of the paper’s launch, simply forward of Apple’s annual Worldwide Builders Convention (WWDC), added to the impression. It shortly went viral throughout X, the place many interpreted the findings as a high-profile admission that current-generation LLMs are nonetheless glorified autocomplete engines, not general-purpose thinkers. This framing, whereas controversial, drove a lot of the preliminary dialogue and debate that adopted.
Critics take purpose on X
Among the many most vocal critics of the Apple paper was ML researcher and X consumer @scaling01 (aka “Lisan al Gaib”), who posted a number of threads dissecting the methodology.
In one broadly shared publish, Lisan argued that the Apple crew conflated token price range failures with reasoning failures, noting that “all fashions may have 0 accuracy with greater than 13 disks just because they can not output that a lot!”
For puzzles like Tower of Hanoi, he emphasised, the output dimension grows exponentially, whereas the LLM context home windows stay fastened, writing “simply because Tower of Hanoi requires exponentially extra steps than the opposite ones, that solely require quadratically or linearly extra steps, doesn’t imply Tower of Hanoi is tougher” and convincingly confirmed that fashions like Claude 3 Sonnet and DeepSeek-R1 usually produced algorithmically appropriate methods in plain textual content or code—but had been nonetheless marked mistaken.
One other publish highlighted that even breaking the duty down into smaller, decomposed steps worsened mannequin efficiency—not as a result of the fashions failed to know, however as a result of they lacked reminiscence of earlier strikes and technique.
“The LLM wants the historical past and a grand technique,” he wrote, suggesting the actual downside was context-window dimension somewhat than reasoning.
I raised one other essential grain of salt myself on X: Apple by no means benchmarked the mannequin efficiency in opposition to human efficiency on the identical duties. “Am I lacking it, or did you not examine LRMs to human perf[ormance] on [the] similar duties?? If not, how have you learnt this similar drop-off in perf doesn’t occur to individuals, too?” I requested the researchers immediately in a thread tagging the paper’s authors. I additionally emailed them about this and plenty of different questions, however they’ve but to reply.
Others echoed that sentiment, noting that human downside solvers additionally falter on lengthy, multistep logic puzzles, particularly with out pen-and-paper instruments or reminiscence aids. With out that baseline, Apple’s declare of a elementary “reasoning collapse” feels ungrounded.
A number of researchers additionally questioned the binary framing of the paper’s title and thesis—drawing a tough line between “sample matching” and “reasoning.”
Alexander Doria aka Pierre-Carl Langlais, an LLM coach at power environment friendly French AI startup Pleias, mentioned the framing misses the nuance, arguing that fashions could be studying partial heuristics somewhat than merely matching patterns.
Ethan Mollick, the AI centered professor at College of Pennsylvania’s Wharton College of Enterprise, known as the concept LLMs are “hitting a wall” untimely, likening it to comparable claims about “mannequin collapse” that didn’t pan out.
In the meantime, critics like @arithmoquine had been extra cynical, suggesting that Apple—behind the curve on LLMs in comparison with rivals like OpenAI and Google—could be making an attempt to decrease expectations,” arising with analysis on “the way it’s all pretend and homosexual and doesn’t matter anyway” they quipped, declaring Apple’s popularity with now poorly performing AI merchandise like Siri.
Briefly, whereas Apple’s examine triggered a significant dialog about analysis rigor, it additionally uncovered a deep rift over how a lot belief to position in metrics when the check itself could be flawed.
A measurement artifact, or a ceiling?
In different phrases, the fashions might have understood the puzzles however ran out of “paper” to write down the total resolution.
“Token limits, not logic, froze the fashions,” wrote Carnegie Mellon researcher Rohan Paul in a broadly shared thread summarizing the follow-up exams.
But not everybody is able to clear LRMs of the cost. Some observers level out that Apple’s examine nonetheless revealed three efficiency regimes — easy duties the place added reasoning hurts, mid-range puzzles the place it helps, and high-complexity instances the place each normal and “considering” fashions crater.
Others view the talk as company positioning, noting that Apple’s personal on-device “Apple Intelligence” fashions path rivals on many public leaderboards.
The rebuttal: “The Phantasm of the Phantasm of Pondering”
In response to Apple’s claims, a brand new paper titled “The Phantasm of the Phantasm of Pondering” was launched on arXiv by unbiased researcher and technical author Alex Lawsen of the nonprofit Open Philanthropy, in collaboration with Anthropic’s Claude Opus 4.
The paper immediately challenges the unique examine’s conclusion that LLMs fail attributable to an inherent incapability to motive at scale. As a substitute, the rebuttal presents proof that the noticed efficiency collapse was largely a by-product of the check setup—not a real restrict of reasoning functionality.
Lawsen and Claude display that lots of the failures within the Apple examine stem from token limitations. For instance, in duties like Tower of Hanoi, the fashions should print exponentially many steps — over 32,000 strikes for simply 15 disks — main them to hit output ceilings.
The rebuttal factors out that Apple’s analysis script penalized these token-overflow outputs as incorrect, even when the fashions adopted an accurate resolution technique internally.
The authors additionally spotlight a number of questionable job constructions within the Apple benchmarks. A number of the River Crossing puzzles, they notice, are mathematically unsolvable as posed, and but mannequin outputs for these instances had been nonetheless scored. This additional calls into query the conclusion that accuracy failures signify cognitive limits somewhat than structural flaws within the experiments.
To check their concept, Lawsen and Claude ran new experiments permitting fashions to provide compressed, programmatic solutions. When requested to output a Lua operate that might generate the Tower of Hanoi resolution—somewhat than writing each step line-by-line—fashions all of the sudden succeeded on way more advanced issues. This shift in format eradicated the collapse solely, suggesting that the fashions didn’t fail to motive. They merely failed to adapt to a man-made and overly strict rubric.
Why it issues for enterprise decision-makers
The back-and-forth underscores a rising consensus: analysis design is now as essential as mannequin design.
Requiring LRMs to enumerate each step might check their printers greater than their planners, whereas compressed codecs, programmatic solutions or exterior scratchpads give a cleaner learn on precise reasoning capacity.
The episode additionally highlights sensible limits builders face as they ship agentic methods—context home windows, output budgets and job formulation could make or break user-visible efficiency.
For enterprise technical determination makers constructing purposes atop reasoning LLMs, this debate is greater than tutorial. It raises crucial questions on the place, when, and the right way to belief these fashions in manufacturing workflows—particularly when duties contain lengthy planning chains or require exact step-by-step output.
If a mannequin seems to “fail” on a posh immediate, the issue might not lie in its reasoning capacity, however in how the duty is framed, how a lot output is required, or how a lot reminiscence the mannequin has entry to. That is significantly related for industries constructing instruments like copilots, autonomous brokers, or decision-support methods, the place each interpretability and job complexity might be excessive.
Understanding the constraints of context home windows, token budgets, and the scoring rubrics utilized in analysis is important for dependable system design. Builders might have to contemplate hybrid options that externalize reminiscence, chunk reasoning steps, or use compressed outputs like capabilities or code as a substitute of full verbal explanations.
Most significantly, the paper’s controversy is a reminder that benchmarking and real-world software usually are not the identical. Enterprise groups needs to be cautious of over-relying on artificial benchmarks that don’t mirror sensible use instances—or that inadvertently constrain the mannequin’s capacity to display what it is aware of.
Finally, the large takeaway for ML researchers is that earlier than proclaiming an AI milestone—or obituary—be sure the check itself isn’t placing the system in a field too small to assume inside.
