Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11The Bitter Lesson offers organizations a reason to take general-purpose AI seriously—not a guarantee that bigger models will solve every business problem. Richard Sutton’s 2019 essay argues, in effect, that AI methods able to make increasing use of computation have historically outperformed systems built mainly around hand-coded human expertise. For adoption, the practical lesson is to test adaptable models against real workflows while measuring errors, oversight costs, security, and business outcomes.
What is the Bitter Lesson?
Richard Sutton’s essay, dated March 13, 2019, draws on the history of AI research in areas including chess, Go, speech recognition, and computer vision. Its central argument is that general methods that can exploit more computation have tended, over time, to be more effective than approaches that depend chiefly on encoding detailed human knowledge. This is a paraphrase of the essay’s argument, not a verified verbatim quotation.
Applied to generative AI, that history warns organizations against assuming that a bespoke system full of hand-written rules will remain the best approach as general-purpose models improve. But the lesson is about a historical pattern in AI research; it is not proof that every model will keep improving indefinitely, or that a general model is suitable for every task.
Why the principle matters now—and where it stops
The International Scientific Report on the Safety of Advanced AI identifies three contributors to recent general-purpose AI progress: more training compute, more training data, and improved training methods. Its 2025 estimates describe approximate annual increases of 4× in training compute, 2.5× in dataset size, and 1.5–3× in algorithmic efficiency. These are estimates of recent trends, not guaranteed forecasts; the report also describes bottlenecks and disagreement over future progress.
#1 Best Overall
The report conditionally projects that, if recent trends continue, some models by the end of 2026 could use 40–100 times the compute of the most compute-intensive models published in 2023, alongside methods using compute 3–20 times more efficiently. That is a conditional forecast, not an observed result. Compute, data, chips, capital, and energy constrain development, and experts disagree about how quickly progress will continue and whether scaling will resolve fundamental challenges such as causal reasoning. [International Scientific Report on the Safety of Advanced AI (2025)]
Capability in a model also does not automatically transfer to a real-world workflow. A 2024 Nature Machine Intelligence editorial notes that, despite high expectations for generative and vision-language models in robotics, real-world complexity remains challenging. That is a robotics-specific caution, not evidence about every industry. [Nature Machine Intelligence editorial (2024)]
Rank #2
What current adoption evidence says—and what it does not
Adoption is substantial in U.S. surveys, but the measures capture different populations and periods. They should not be combined as though they estimate the same thing.
| Evidence | Population and period | Reported finding | How to interpret it |
|---|---|---|---|
| Bick, Blandin, and Deming, Management Science (2026) | U.S. residents ages 18–64 and employed respondents; survey findings through late 2024 | 45% of residents ages 18–64 used generative AI; 27% of employed respondents used it for work in the prior week, including 10% every workday. Respondents’ reported time savings equated to 1.4% of total work hours. | These are survey measures and respondent-stated time savings, not causal proof of productivity gains from a particular organization’s rollout. The authors report variation by industry and note the importance of workplace climate and policies. |
| U.S. Census Bureau Center for Economic Studies (2026) | U.S. firms; survey supplement covering November 2025–January 2026 | 18% of firms used AI in a business function, or 32% when weighted by employment. Among adopters, 57% used AI in three or fewer functions; 65% of firms limited AI use to three or fewer tasks. | The paper distinguishes firm adoption, functions, and worker tasks. Writing, document analysis, and information search were leading worker-task uses. It finds a positive correlation between broader integration and commercial performance, not proof that integration caused better performance. |
The Bick, Blandin, and Deming study reports that work adoption of generative AI was faster than PC adoption relative to each technology’s first mass-market launch. That comparison describes adoption pace, not equivalent economic effects. The Census paper also finds diffusion from both directions: workers may use AI without formal firm adoption, and firms may formally adopt AI without broad worker-task use. Most measured users relied on AI to augment tasks, while AI-related employment decreases were rare in the paper’s measures. Neither finding establishes what will happen at a specific firm. [Bick, Blandin, and Deming, Management Science (2026)] [U.S. Census Bureau Center for Economic Studies working paper (2026)]
How to apply the lesson to an adoption decision
Use the Bitter Lesson as a reason to keep general-purpose approaches in consideration, not as a substitute for evaluating a specific workflow. Compare candidate approaches on the dimensions that determine whether deployment is useful and safe.
| Decision axis | What to assess |
|---|---|
| Task fit and error cost | Test representative inputs from the actual workflow. Identify what a wrong answer would do, how often it can be detected, and who is accountable for fixing it. |
| Human oversight and security | Decide which outputs require review and test for risks such as prompt injection, jailbreaks, and data poisoning. Match safeguards to the consequences of failure. |
| Adoption depth | Measure individual task use, deployment across business functions, and integration into operational workflows separately; one does not imply the others. |
| Organizational readiness | Plan staff engagement, training and support, risk management, and ongoing monitoring before treating a tool as an established part of work. |
| Evidence of value | Measure local outcomes against a relevant baseline. Do not infer a business result from a general benchmark, survey-reported time savings, or another organization’s adoption level. |
Build evaluation and oversight into deployment
The U.S. Government Accountability Office describes practices including benchmark testing, multidisciplinary evaluation, and red-teaming. It also notes that generative AI can produce incorrect or biased outputs and can be vulnerable to prompt injection, jailbreaks, and data poisoning. Public disclosure of training-data specifics is limited, so organizations should not assume they can fully inspect how a model was built.
GAO’s practical point is that “user judgment should play a role in accepting model outputs.” How much review is appropriate depends on the task and the harm a mistake could cause: a low-consequence drafting aid and a system influencing a consequential decision do not call for identical controls. [U.S. Government Accountability Office technology assessment (October 22, 2024)]
For implementation, UK government guidance organizes its People Factor and Mitigating Hidden AI Risks Toolkit around “Adopt, Sustain, Optimise.” It is intended for people involved in AI development, delivery, procurement, and governance, and covers engagement, training and support, risk management, and monitoring. It is guidance for organizing adoption work, not a guarantee of success. [UK government AI adoption toolkit]
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsMake the decision from local evidence
The Bitter Lesson is a useful correction to the instinct that human expertise, encoded in custom rules, will always be the strongest foundation for an AI system. Its counterpart is just as important: a scalable general method is not automatically reliable, economical, secure, or valuable in a particular setting. Choose by testing the work that needs doing, setting appropriate human and technical controls, and tracking outcomes at the level where adoption actually occurs.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
