A PPT Dismissed as 'Nonsense' by MIT Professors 5 Years Ago Predicted the Core Ideas of OpenAI o1 and o3

marsbit2026-08-17 tarihinde yayınlandı2026-08-17 tarihinde güncellendi

Özet

In 2020, AI researcher Giambattista Parascandolo presented his vision for neural network reasoning at an MIT faculty interview, only to have the committee dismiss the direction as "nonsense." He later posted the details online. His talk centered on enabling artificial neural networks to generalize and plan beyond their training data, closer to human capabilities. Parascandolo proposed three key research directions. First was "open-ended reasoning," where models could dedicate more computation time to harder problems, continuously refining answers—a precursor to today's compute-adaptive reasoning models. He noted that simply adding steps (e.g., in RNNs) wasn't enough without learning to use them effectively. Second, he advocated using language as a medium for reasoning within reinforcement learning. By leveraging the world knowledge in models like GPT, agents could better describe environments, decompose tasks, and plan—foreshadowing concepts like chain-of-thought and agent workflows. His third direction involved giving AI systems the ability to manipulate their own learning process: resetting to past states, creating counterfactual scenarios, and even editing their own activations or weights to facilitate deliberate practice. Parascandolo, who earned his PhD focusing on out-of-distribution generalization and had internships at Google X and DeepMind, joined OpenAI in 2021. He contributed to GPT-4 and later became integral to the foundational research behind the reasoning mod...

The AI community never runs short of gossip.

This time, the protagonist is Giambattista Parascandolo, an important researcher in the direction of reasoning models at OpenAI.

In 2020, he went to MIT to interview for a professorship, giving a presentation on using GPT for reasoning. As a result, most professors on the interview committee dismissed this direction as "nonsense."

https://x.com/turingbook/status/2084003612687249836?s=20

This guy went all out, directly putting this experience on his personal homepage and attaching the description and slide link from that year's presentation.

https://sites.google.com/view/giambattista-parascandolo/home

What was in that report dismissed as "nonsense"?

The MIT website shows that the topic of this presentation was how to enable artificial neural networks to break through the training distribution and achieve generalization and planning capabilities closer to humans.

Parascandolo believed that humans can recombine existing knowledge, identify key invariants, build abstract models, and perform long-horizon planning. Artificial neural networks still have much room for improvement in these aspects.

At the end of the presentation, he proposed three future research directions: open-ended reasoning in neural networks, unexplored degrees of freedom in artificial neural networks, and using language as a reasoning substrate in reinforcement learning to improve sample efficiency.

The most crucial concept among these was "open-ended reasoning."

Parascandolo defined it as: the model can invest more time and computation to continuously refine its answer. The harder the problem, the more steps the model should think, and the extra computation should translate into better results.

This already sounds very much like today's inference-time compute scaling.

At that time, standard Transformers had a fixed network depth, and the forward compute each token experienced was basically predetermined. However, problem difficulty has no stable relationship with input length. A problem could be very long but have a simple answer. Another problem might be a single sentence but require multiple rounds of decomposition and verification.

RNNs seemed more suitable for such tasks. They could run repeatedly and theoretically obtain arbitrary amounts of thinking time. However, the curve Parascandolo showed in his PPT indicated that RNNs typically performed best only around the number of reasoning steps they had seen during training. Continuing to increase the number of iterations could actually decrease accuracy.

This meant that increasing compute was only the first step; the model also needed to learn how to utilize this compute.

At the time, the most effective multi-step planning heavily relied on model-predictive control and Monte Carlo tree search. Neural networks were responsible for predicting the environment or evaluating value, while external search algorithms were responsible for expanding future paths. Parascandolo hoped to further integrate long-horizon reasoning capabilities into neural networks themselves.

His second vision was to make language the substrate for reasoning.

Parascandolo used the classic game *Montezuma's Revenge* as an example. A reinforcement learning Agent trained from scratch needed to try a huge number of state and action combinations. Many correct actions themselves were not complex; what the Agent truly lacked was a judgment of "which behaviors are more reasonable."

GPT had already absorbed vast amounts of world knowledge from text. Language could help models describe environments, understand goals, decompose tasks, and generate high-level plans, also significantly narrowing the search space.

Today, this line of thinking easily brings to mind chain-of-thought, language planning, and Agent workflows.

The third direction Parascandolo proposed was that AI systems could reset tasks, return to any remembered state, construct counterfactual scenarios, and even adjust time, gravity, and observations within the simulator. Systems could also directly read, copy, and modify their own activation values and neural network weights.

This was equivalent to incorporating the learning process itself into the Agent's operational space. It could engage in deliberate practice, generate special training scenarios, transfer existing knowledge, and relearn from failure trajectories.

The PPT from five years ago almost wrote today's reasoning model roadmap.

We also dug up a blog post he wrote in June 2021, titled "Backprop, Evolution, and the ‘Two Dogs’ Fallacy."

This blog post mainly refuted the view that "neural networks need massive amounts of data, therefore they are unlike the human brain."

Parascandolo argued that the fact humans can learn to recognize dogs after seeing just a few does not mean there was no prior accumulation of experience. Long evolution has already condensed the experiences of ancestors interacting with the world into the structure and inductive biases of the human brain.

From this perspective, the large-scale pre-training of neural networks can be analogized to biological evolution, while model fine-tuning and in-context learning are closer to the learning an individual does in a lifetime. GPT-3 has read far more text than any individual, but its pre-training served to compress massive experience and shape efficient learning capabilities. Therefore, one should not directly compare the model's entire pre-training data with the small number of learning samples an individual receives after birth.

This understanding also implies that continuing to scale data and compute could still bring significant improvements. He was drawing an analogy about the roles they play, not claiming that gradient descent and biological evolution share the same specific mechanisms.

What's this guy's background?

This guy is quite low-key. His latest post on X dates back to November 2024, when he was recruiting two Research Engineers (RE) and Software Engineers (SWE) for o1.

According to his personal homepage, Parascandolo's research experience has always revolved around generalization, planning, and reasoning.

In 2017, he began his Ph.D. at the Max Planck Institute for Intelligent Systems and ETH Zurich, supervised by Bernhard Schölkopf and Thomas Hofmann. His doctoral research focused on OOD generalization in deep learning.

During his Ph.D., he did internships at Google X in Mountain View and DeepMind in London, participating in research on automated design on ultra-large-scale simulators at the former, and research on divide-and-conquer Monte Carlo tree search at the latter.

After graduating with his Ph.D. in September 2021, he directly joined OpenAI. Initially, he entered the reinforcement learning team led by John Schulman, then moved to the algorithms team where Mark Chen was, and later joined the 🍓 (strawberry) team led by Jerry Tworek. The strawberry team was the internal codename for the o1 project.

In 2023, he participated in GPT-4 development and formed a new team continuing reasoning research. Later, he participated in the foundational research for OpenAI o1 and o3, promoting the scaling of reward modeling, environment construction, and general reasoning algorithms. Part of the algorithm description is still covered by black blocks on his page.

Parascandolo didn't get the MIT professorship from that 2020 interview; he went to OpenAI instead.

Four years later, he helped build o1.

Life is always full of delightful surprises.

References:

https://x.com/giambattista92

https://sites.google.com/view/giambattista-parascandolo/home

https://gibipara92.github.io/2021/06/09/backprop-evolution-two-dogs.html

https://docs.google.com/presentation/d/1edd2GkAP31wNygaev3DO8gE1t2lgsx1z8TeVpBKqlYA/edit?slide=id.gc772610149_0_179#slide=id.gc772610149_0_179

This article is from the WeChat public account "Almost Human" (ID: almosthuman2014), author: Yang Wen

Trend Kriptolar

İlgili Sorular

QWhat were the three key research directions proposed by Giambattista Parascandolo in his MIT job talk?

AThe three key research directions were: 1) Open-ended reasoning in neural networks (models allocating more time/computation for harder problems), 2) Using language as a substrate for reasoning in reinforcement learning to improve sample efficiency, and 3) Allowing AI systems to manipulate their own internal states (like activations and weights) to conduct deliberate practice and counterfactual learning.

QHow does Parascandolo's concept of 'open-ended reasoning' relate to modern AI like OpenAI's o1?

AParascandolo's 'open-ended reasoning' concept, where models spend more time and computation to refine answers for harder problems, directly foreshadows the core idea of 'reasoning-time compute scaling' used in modern AI systems like OpenAI's o1 and o3. These models dynamically allocate computational effort based on problem difficulty.

QWhy did Parascandolo argue that comparing a neural network's pre-training data to a human's lifelong learning is misleading?

AParascandolo argued that large-scale pre-training is analogous to biological evolution, compressing vast ancestral experience into the model's structure and inductive biases. Fine-tuning and in-context learning are more akin to an individual's lifetime learning. Therefore, comparing total pre-training data to a human's post-birth learning samples is an apples-to-oranges comparison.

QWhat role did Giambattista Parascandolo play at OpenAI leading up to the development of o1 and o3?

AAt OpenAI, Parascandolo initially worked in the reinforcement learning team, later joined the algorithms team, and was part of the 'Strawberry' team (the internal codename for the o1 project). He contributed to GPT-4, formed a new team focused on reasoning research, and worked on the foundational research for o1 and o3, specifically on reward modeling, environment construction, and scaling general reasoning algorithms.

QWhat was the main point of Parascandolo's blog post 'Backprop, Evolution and the "Two Dogs" Fallacy'?

AThe main point was to challenge the argument that neural networks are unlike the human brain because they require massive data. He posited that human ability to learn from few examples (like recognizing dogs) is built upon evolutionary history, just as a neural network's pre-training compresses vast experience. Thus, pre-training is analogous to evolution, not individual learning, justifying the value of scaling data and compute.

İlgili Okumalar

Wall Street Morning Report: S&P Earnings Growth Hits 30-Year High, Nvidia Becomes SpaceX's 6th Largest Shareholder

Wall Street Morning Report: S&P 500 earnings growth hits a 30-year high; Nvidia becomes SpaceX's 6th largest shareholder. Market Summary: U.S. stocks ended slightly lower on Friday after record highs, with the S&P 500 posting a weekly gain. Weaker-than-expected July retail sales data significantly reduced market expectations for a September Fed rate hike. Key Themes: * **Macro & Rates:** The probability of the Fed holding rates steady in September rose to ~70%. Geopolitical tensions in the Middle East supported oil prices near $90. Focus this week is on a $16B 20-year Treasury auction testing investor demand amid high yields. * **Corporate Earnings:** S&P 500 Q2 earnings grew 31%, the fastest pace since 1992 excluding recession recoveries, driving a valuation reset. * **AI & Tech Sector Rotation:** The AI investment narrative faces scrutiny over financing and return timelines. Semiconductor stocks saw divergence: memory (SanDisk, Micron) and optical communication (Applied Optoelectronics) surged, while Broadcom fell sharply on credit rating downgrade concerns over AI leasing exposure. * **Notable Moves:** Nvidia disclosed a ~$21B stake in SpaceX. AMD rose 6.5% after a record debt issuance. Reddit rallied ahead of its S&P 500 inclusion. This Week's Highlights: * **Fed Minutes (Thu):** Key for gauging policy divergence and September rate outlook. * **Key Earnings:** Alibaba, Walmart, Nvidia, AMD. * **Events:** Seoul AI Summit, World Robot Conference, potential恒生指数 rebalancing. * **Data:** U.S. jobless claims, Japan core CPI, Korea early export data (a leading indicator for semiconductors).

marsbit23 dk önce

Wall Street Morning Report: S&P Earnings Growth Hits 30-Year High, Nvidia Becomes SpaceX's 6th Largest Shareholder

marsbit23 dk önce

BTC in Range-Bound Consolidation, HYPE Daily Rebound Confirmed | Invited Analysis

**Market Analysis: BTC Consolidates in Range, HYPE Shows Daily Rebound Confirmation (Weekly Outlook)** **BTC Analysis:** The market is at a key juncture. Bitcoin is undergoing a daily correction, with a critical focus on whether it can find support and stabilize above the $60,950 level. The adjustment that began from the May 6th high has formed a complex seven-segment structure on the daily chart. Three potential paths are identified: 1. **Continuation of the c-wave rally** (primary scenario), targeting $67,300 initially, then $69,500-$71,000. 2. **Range-bound consolidation** between $60,950 and $65,500, forming an "ascending中枢" before an upward breakout. 3. **Breakdown and secondary探底** if support at $60,950-$61,500 fails. Analysis favors the first two scenarios. The 4-hour chart suggests the current adjustment phase is nearing completion near the strong $60,950-$61,500 support zone. **BTC Strategy:** * **Key Levels:** Support at $62,268, $60,950-$61,500, $57,820. Resistance at $65,500, $67,300, $69,500-$71,000. * **Mid-term:** Current model indicates a bearish structure with ~20% short exposure. Consider adding to shorts (up to 50%) only if price rallies to $69,500-$71,000, shows clear stagnation, and model confirms. * **Short-term (30% capital):** Two tactical plans: * **A) Test shorts** near strong resistance ($69,500-$71,000) upon signs of weakness. * **B) Test longs** near strong support ($60,950-$61,500) upon stabilization signals. **HYPE Analysis:** HYPE has established a daily rebound from the August 2nd low. The 4-hour chart shows a five-segment上升 structure with a formed "ascending中枢". The current (76-77) segment is crucial. Two outcomes are possible: 1. **Breakout (Central离开段):** If price breaks and holds above the $58-$58.5 resistance zone, the uptrend continues. Monitor for momentum背驰 between the initial and breakout segments. 2. **Consolidation (Central Extension):** Failure to hold above $58-$58.5 leads to a return to中枢 consolidation, requiring new signals for direction. **HYPE Strategy:** * **Key Levels:** Support at $50-$52, $45. Resistance at $58.5-$60, descending channel upper rail, $72.97. * **Watch:** Whether price can break/hold above $58-$58.5 to confirm the breakout. * **Action:** * **Existing longs (from $50-$52):** Move stop-loss above entry cost. * **New entrants:** Consider light longs only on a confirmed breakout above $58.5-$60 with a tight stop-loss. Exit quickly if momentum背驰 appears. **General Risk Management:** * Set initial stop-loss immediately upon entry. * Move stop-loss to breakeven at +1% profit. * Trail stop-loss upwards by 1% for every subsequent 1% gain to lock in profits. **Disclaimer:** Market conditions change rapidly. This analysis is for informational purposes only and not investment advice. Trading carries significant risk.

marsbit26 dk önce

BTC in Range-Bound Consolidation, HYPE Daily Rebound Confirmed | Invited Analysis

marsbit26 dk önce

Not Chasing AI or Buying Back Shares, Can 'Stingy' Jingdong Still Succeed?

On August 13th, JD.com (JD.US) released its Q2 2026 earnings, delivering a mixed and generally "flat" performance that met low expectations but offered few positives. Overall revenue declined by approximately 3% year-over-year (YoY) to ~¥346.4 billion, aligning with weakened domestic consumption trends. While group operating profit saw a YoY improvement, this was primarily due to reduced losses in the food delivery segment compared to the high-cost "delivery war" period of the previous year. Key details reveal deeper concerns: revenue from JD's core domestic retail segment fell 4.7% YoY. While electronics sales declined less than feared (~12% YoY), growth in daily necessities and advertising services plummeted by about 10 percentage points each, raising doubts about the segment's mid-term growth momentum post-subsidy adjustments. Logistics revenue growth also slowed to 5.9% as the delivery boom faded. Profitability presented a nuanced picture. The retail segment's operating margin edged up slightly YoY but failed to deliver the significant beats seen in past quarters, suggesting efficiency gains may be nearing limits. Losses from the New Businesses segment (including food delivery and overseas ventures) remained elevated at ~¥9.9 billion, as increased overseas investment partially offset reductions in delivery subsidies. Notably, JD's shareholder returns weakened significantly, with share buybacks in H1 2026 annualizing to only about 5% of market cap. The company opted to park cash in short-term investments rather than boost returns, drawing criticism. Looking ahead, JD's performance hinges on a potential recovery in China's e-commerce sentiment in H2 2026 and the scale of ongoing losses in new ventures. While not burdened by massive AI capex like some peers, and offering relative defensive stability, the lack of positive earnings surprises and diminished shareholder returns provide little compelling reason for investors to favor the stock in the near term.

marsbit36 dk önce

Not Chasing AI or Buying Back Shares, Can 'Stingy' Jingdong Still Succeed?

marsbit36 dk önce

İşlemler

Spot

Popüler Makaleler

CORE Nasıl Satın Alınır

HTX.com’a hoş geldiniz! CORE (CORE) satın alma işlemlerini basit ve kullanışlı bir hâle getirdik. Adım adım açıkladığımız rehberimizi takip ederek kripto yolculuğunuza başlayın. 1. Adım: HTX Hesabınızı OluşturunHTX'te ücretsiz bir hesap açmak için e-posta adresinizi veya telefon numaranızı kullanın. Sorunsuzca kaydolun ve tüm özelliklerin kilidini açın. Hesabımı Aç2. Adım: Kripto Satın Al Bölümüne Gidin ve Ödeme Yönteminizi SeçinKredi/Banka Kartı: Visa veya Mastercard'ınızı kullanarak anında CORE (CORE) satın alın.Bakiye: Sorunsuz bir şekilde işlem yapmak için HTX hesap bakiyenizdeki fonları kullanın.Üçüncü Taraflar: Kullanımı kolaylaştırmak için Google Pay ve Apple Pay gibi popüler ödeme yöntemlerini ekledik.P2P: HTX'teki diğer kullanıcılarla doğrudan işlem yapın.Borsa Dışı (OTC): Yatırımcılar için kişiye özel hizmetler ve rekabetçi döviz kurları sunuyoruz.3. Adım: CORE (CORE) Varlıklarınızı SaklayınCORE (CORE) satın aldıktan sonra HTX hesabınızda saklayın. Alternatif olarak, blok zinciri transferi yoluyla başka bir yere gönderebilir veya diğer kripto para birimlerini takas etmek için kullanabilirsiniz.4. Adım: CORE (CORE) Varlıklarınızla İşlem YapınHTX'in spot piyasasında CORE (CORE) ile kolayca işlemler yapın.Hesabınıza erişin, işlem çiftinizi seçin, işlemlerinizi gerçekleştirin ve gerçek zamanlı olarak izleyin. Hem yeni başlayanlar hem de deneyimli yatırımcılar için kullanıcı dostu bir deneyim sunuyoruz.

417 Toplam GörüntülenmeYayınlanma 2024.12.13Güncellenme 2026.06.02

CORE Nasıl Satın Alınır

Tartışmalar

HTX Topluluğuna hoş geldiniz. Burada, en son platform gelişmeleri hakkında bilgi sahibi olabilir ve profesyonel piyasa görüşlerine erişebilirsiniz. Kullanıcıların CORE (CORE) fiyatı hakkındaki görüşleri aşağıda sunulmaktadır.

活动图片