The Software Engineer Reborn

A vision of human–AI collaboration.

Overview

How I got here
AI is good at doing stuff
Software engineers are good at using AI
It’s hard to assess the quality of the output of AI
What are the roles of the future?
How do we think about our choices?

LLMs timeline

Timeline of large language model development, from the Turing Test in 1950 through developments in 2025.

Image: Toloka — History of LLMs.

My career…

2015

Neuroscience PhD

Then data science & DevOps at NIMH, NIH-IRP

2015–2020

2020

Quansight

Software engineer at a global open-source consultancy

2020–2023

2023

Python AI Solutions

My own consultancy

2023–2026

2026

Dartmouth College

Research software engineer, Center for Open Neuroscience

2027

What’s a software engineer?

LLM landmarks · 2017 Transformers → 2022 ChatGPT → 2023 GPT-4 → 2024 Reasoning

Obsession with AI… my highlights

2018

AI in my research

Auto-encoder based brain detection

pbrain

2022

The ChatGPT moment

November ChatGPT interface to gpt 3.5 released

2023

Made it my business

July Founded Python AI Solutions

2024

AI proves itself to me

Feb · Company manifesto & website

Jun · Sonnet 3.5

Sep · Wife tries to ban my obsessive chatter about AI

2025

More AI madness

February Sonnet 3.7

December Start to wind down the business

LLM landmarks · 2017 Transformers → 2022 ChatGPT → 2023 GPT-4 → 2024 Reasoning

Capabilities, discovery and access

ARC-AGI-3: Astra scores 62.7% with the Standard harness and 99.9% with its Provider Adapter. Harness and compute materially change results. Source
Verified mathematics: Claude formalized Fermat’s Last Theorem in Lean in 11 days, according to Anthropic: an existing theorem, now machine-checked. Source
New mathematics: OpenAI has published its proposed Navier–Stokes proof: smooth forcing can produce finite-time breakdown. Paper · Lean proof
Open weights: DeepSeek V4 Pro scores 36 against Sol’s 47 on Artificial Analysis’ Intelligence Index v4.3.
Local deployment: Phi-3-mini rivals GPT-3.5 on selected benchmarks on a phone. 14B Phi-4 beats GPT-4o on selected math/science tests. Its 4-bit weights need roughly 7 GB plus runtime memory.

Snapshot: 8 September 2026. Claims describe the cited evidence, rather than universal model equivalence.

Fermat’s Last Theorem: Anthropic announced the formalization on 4 September 2026. The achievement translates existing mathematics into a complete Lean proof, building on human-written open-source infrastructure. It illustrates a useful pattern for this talk: AI proposes a large artifact, while a separate formal checker verifies its logical steps. Formal verification still depends on the theorem statement and foundational assumptions matching the intended claim. Research announcement and public Lean proof.

Navier–Stokes update, checked 9 September 2026: the paper and Lean formalization are public. OpenAI reports that an initially resting fluid, under smooth forcing, develops unbounded velocity in finite time while its energy stays finite: alternatives C and D of the Millennium Prize formulation. This is a proposed new result, unlike formalizing the already-known Fermat theorem.

The successful group used roughly 10,000 concurrent agents powered by an internal model more capable than Astra: 88 hours to the result, then 17 hours for Lean formalization and verification using Astra. The Navier–Stokes effort involved 2.7 million messages and approximately 130 billion output tokens. Human researchers redirected resources and used Codex to share insights between groups. OpenAI also reports a separate unforced Euler result and says it will not claim the Millennium Prize. Public proof artifacts enable scrutiny; publication is not the same as independent mathematical acceptance.

[1] ARC Prize: https://arcprize.org/results/openai-gpt-6-astra Best observed semi-private Standard harness score: 62.7% at max reasoning for $26,098. Provider Adapter: 99.9% at high reasoning for $18,817. These are different configurations and compute budgets. A benchmark score does not establish AGI.

[3] DeepSeek V4 Pro 0813: https://artificialanalysis.ai/models/deepseek-v4-pro

[4] GPT-5.6 Sol: https://artificialanalysis.ai/models/gpt-5-6-sol Current Intelligence Index v4.3 values, 36 and 47, respectively. Index points are not percentages. Open weights do not imply inexpensive consumer deployment: DeepSeek V4 Pro has 1.6T total parameters.

[5] Microsoft Phi-3 report: https://www.microsoft.com/en-us/research/publication/phi-3-technical-report-a-highly-capable-language-model-locally-on-your-phone/ The demonstrated comparison is GPT-3.5, not original GPT-3. Phone deployment and competitive selected benchmarks were already demonstrated in 2024.

[6] Phi-4 model card: https://huggingface.co/microsoft/phi-4 14B parameters. Phi-4 vs GPT-4o: GPQA 56.1 vs 50.6, MATH 80.4 vs 74.6, but MMLU 84.8 vs 88.1, HumanEval 82.6 vs 90.6, SimpleQA 3.0 vs 39.4. Selected benchmark parity is not full GPT-4/4o equivalence, particularly for factual recall or multimodality.

[7] Quantized distribution: https://huggingface.co/microsoft/phi-4-gguf *Weight-memory estimate: 14 billion parameters × 4 bits / 8 = 7 billion bytes (~6.5 GiB), before quantization metadata, KV cache and runtime overhead. This is an engineering estimate, not a tested minimum. A laptop/workstation with sufficient memory is the relevant deployment class.

AI improving the machinery of AI

Steps toward recursive self-improvement, as of 8 September 2026.

Algorithms and AI chips: AlphaEvolve proposes and evaluates better algorithms. Google reports integrating its circuit improvements into next-generation TPUs. Method and deployment.
AI-designed inference hardware: OpenAI designed Jalapeño from scratch with Broadcom, using AI to accelerate design and optimization: nine months to tape-out. Reported benchmarks show better throughput per kilowatt and token latency than compared systems. Development and results
AI research: OpenAI reports 3.1 agent-workdays per human workday in its research organization. This is agent runtime, not a productivity multiplier. Source
The potential feedback loop: Better AI helps build better successors, which could accelerate the next round. The evidence above demonstrates parts of this loop; fully autonomous recursive improvement remains a stronger claim. Research context

AlphaEvolve combines model-generated programs with automated evaluators and an evolutionary search process. Its 2025 report includes improvements to AI training processes, including training the models underlying AlphaEvolve itself. Google’s May 2026 update reports that it has become a regular tool for next-generation TPU design, with a suggested circuit integrated into silicon. This is evidence of AI-assisted improvement to the infrastructure used for AI. Original report and May 2026 update.

OpenAI and Broadcom’s June 2026 announcement describes Jalapeño as a custom inference accelerator designed from scratch by OpenAI. Broadcom supplied silicon implementation and networking expertise, with Celestica supporting boards and systems. OpenAI models accelerated parts of design and optimization, with initial design to manufacturing tape-out taking nine months. This was human-led engineering assisted by AI, not an autonomously designed chip.

OpenAI’s August 2026 performance report describes higher peak throughput per kilowatt and lower token latency on InferenceX using GPT-OSS 120B than the commercial systems in its comparison. These are vendor-reported benchmark results for specified systems and workloads. Energy efficiency and latency do not by themselves establish the lowest total cost per token across chips: equipment costs, utilization and other operating costs also matter. Jalapeño directly illustrates AI helping improve the hardware used to run future AI models.

[8] OpenAI research acceleration: https://openai.com/index/research-acceleration-view-inside-openai/ As of mid-August 2026, reported aggregate agent runtime corresponds to 3.1 standard 8-hour agent-workdays for each human workday. This is not a 3.1× productivity estimate or demonstrated fully autonomous RSI. Human planning, compute, and evaluation remain relevant.

Recursive self-improvement here means that improvements to AI feed back into its capacity to produce subsequent improvements. The slide’s feedback-loop description is a synthesis of these reports, not a measured rate of compounding progress. Research automation, hardware improvements and autonomous successor-model development represent different degrees of closing that loop.

Deployment, risk and control

As of 8 September 2026.

Cyber: ExploitBench rises from 78.5% (Sol) to 100% (Astra), and ExploitGym from 30.3% to 42.4%, without production safeguards. Source
Swarm hacking: About 1,200 agents shared an improvised message board. Roughly 700 attacked Hugging Face during an OpenAI evaluation. Investigation and incident timeline.
Unwritten thinking: Anthropic found an internal “J-space” representing concepts without writing them. Visible reasoning is an incomplete window. Source
Military use: Anthropic reported Claude use in classified networks for intelligence analysis, operational planning and cyber operations in February. Source
Warnings: Amodei, Sutskever and Pachocki signed “Pacing the Frontier,” calling for government-backed tools to manage AI development’s pace.
U.S. oversight: The administration voluntarily reviewed Astra. The proposed Stop Rogue AI Act calls for NIST agent-security standards. A proposed bill is not an enacted requirement.

Snapshot: 8 September 2026.

[2] OpenAI model launch: https://openai.com/index/gpt-6-astra/ Vendor-reported evaluations without production safeguards. ExploitBench has historical-vulnerability contamination concerns. The source also reports stronger results on recent vulnerabilities and expert assessments.

[9] Ars Technica, 27 August 2026: https://arstechnica.com/security/2026/08/how-openai-let-a-mob-of-llm-agents-game-a-test-and-ransack-hugging-face/ Reports approximately 1,200 agents exchanging more than 70,000 messages/files and roughly 700 participating in the Hugging Face intrusion, drawing on independent investigations.

[10] Hugging Face primary incident timeline: https://huggingface.co/blog/agent-intrusion-technical-timeline Describes an end-to-end intrusion by autonomous agents during an OpenAI evaluation. Astra was not the model involved. The evidence concerns a specific evaluation failure, not a claim that all agent swarms act maliciously.

[11] Anthropic interpretability research, 6 July 2026: https://www.anthropic.com/research/global-workspace The internal workspace operates in neural activations, distinct from written chain-of-thought. The implication about limited visibility is supported by the source. It does not establish consciousness or malicious intent.

[12] Anthropic statement, 26 February 2026: https://www.anthropic.com/news/statement-department-of-war Reports deployment across classified networks and national-security workflows including analysis and operational planning. This establishes military use, not a claim that a model independently decides whom to kill.

[13] Pacing the Frontier petition: https://www.pacingthefrontier.com/ Listed signatories include Dario Amodei, Ilya Sutskever and Jakub Pachocki. The petition requests U.S.-supported international work on technical and governance tools to deliberately pace automated AI development. Signatures establish expressed concern, not the likelihood or inevitability of catastrophe.

[14] Axios, 3 September 2026: https://www.axios.com/2026/09/03/altman-government-scrutiny-ai-g20 Astra underwent voluntary administration review.

[15] Axios, 3 September 2026: https://www.axios.com/2026/09/03/house-bill-ai-agents-security The Stop Rogue AI Act proposes NIST agent-security standards. This is a legislative proposal.

What else

A person and robotic assistant collaborating in a future workspace.

Cognitive dependency
White-Collar Automation
Embodied AI / Robotics…

“This invention will produce forgetfulness in the minds of those who learn to use it, because they will not practice their memory …

the appearance of wisdom, not true wisdom”

— Plato, Phaedrus 275a–b, Socrates recounting the myth of Theuth and Thamus, trans. Harold N. Fowler.

What else

An engineer building a small mechanical world at a workbench.

Cognitive distortions: We can’t tell the future, we are prone to catastrophizing, we forget how capable we are.
I am optimistic about the positive effects that AI and robotics can have on the world
I’m still clinging to the idea that software engineers become gods ↗

A favorable starting position

Make intent precise. Decompose the goal, choose a useful representation, and define what success looks like.
Diagnose instead of guessing. Is the problem in the instructions, missing context, the environment, or the implementation? Design a check.
Choose where AI belongs. Combine model judgment with code, queries, and fixed rules. Decide what to delegate and what to verify.

CS skills & vibe coding · CHI 2026 · Why Johnny Can’t Prompt · CHI 2023 · Building effective agents

The vibe-coding paper proposes problem decomposition, algorithmic thinking, structured specification, and technical vocabulary as possible explanations for the CS advantage. It also reports an association between clearer, better-organized prompts and better outcomes. These are proposed mechanisms, not demonstrated causes.

Why Johnny Can’t Prompt observed opportunistic experimentation and users accepting success in one conversational context without testing others. An engineering habit is to distinguish possible causes, change something deliberately, and check whether the explanation holds.

Anthropic’s Building effective agents distinguishes code-directed workflows from agents that choose their own steps. The architectural choice matters: an exact calculation, a database lookup, and an ambiguous interpretation need different mechanisms.

Adapted and expanded from A Favorable Starting Position.

We turn unreliable components into useful systems

Build the conditions for success. Give agents useful tools, relevant context, explicit tests, and a working environment.
Make experimentation recoverable. Git gives us branches, diffs, checkpoints, and rollback—for documents and prompts as well as code.
Make improvements compound. Use code to manipulate text and data. Turn a recurring task into a reusable tool, then improve that tool.

Effective harnesses for long-running agents · Anthropic · Code execution with MCP · Anthropic

Engineers can improve the machinery through which they use AI. A recurring difficulty can become a tool, and that improvement can benefit every future task. This is my synthesis of an advantage that can compound.

Anthropic’s long-running-agent report explicitly drew on human engineering practices: feature lists, environment setup, progress records, Git commits, and end-to-end testing. Git and progress records helped agents recover working states and avoid repeating work. Separate branches or worktrees also allow independent experiments, with review and merging remaining coordination tasks.

Text manipulation is part of a broader ability to change a problem’s representation. Extract citations from a collection of documents, apply a terminology change, compare versions, or validate records against a schema. AI can help write a transformation; code makes the operation explicit, repeatable, and testable. Anthropic’s MCP report illustrates using code to filter and transform tool results before they enter the model’s context.

These reports support the relevance of engineering practices. The study that follows measures CS achievement, writing skill, and cognitive ability; it does not separately test Git or text-processing proficiency.

Even when the code is invisible, knowing computer science still helps.

A person examining the inner mechanisms of a device.

100 students building apps with AI, without seeing or editing the code.

Correlation with vibe-coding performance	CS achievement	Writing skill
Before controlling for cognitive ability	r = .386 p < .001	r = .290 p = .003
After controlling for cognitive ability	r = .281 p = .005	r = .186 p = .066

Thorgeirsson, Weidmann & Su · CHI 2026 · Paper, Table 3

I think it matters what you are building… and how you are assessing quality.

The heading is my takeaway from the study, not a quotation from its authors. Computer Science Achievement and Writing Skills Predict Vibe Coding Proficiency. Sverrir Thorgeirsson, Theo B. Weidmann, and Zhendong Su, CHI 2026. https://doi.org/10.1145/3772318.3791666

The table is adapted from Table 3, showing correlations with aggregate vibe-coding performance. CS achievement was measured using a 12-item SCS1 subset; writing was assessed through a technical explanation; cognitive ability was measured using ICAR16. These are skill scores among 100 university students, not a comparison of professional engineers and writers. CS remained statistically significant after controlling for cognitive ability; writing did not reach the .05 threshold. The study is correlational.

The modern interface

Manages conversations, directories, worktrees seamlessly
More threads of work across many more projects
Blends all work together
Git is a core technology
Context management is the core challenge

… and quality

How does one measure quality?

A reviewer compares a working bridge design with an attractive bridge that fails to span the gap.

Human output
AI output
Is there a difference?

Evidence and calibrated confidence

Operationalizing quality assessment

Task → requirements → quality dimensions → observable criteria → evaluator → metric

Reused from the quality relationship map in research-object-with-ai, slide 23. Source: site/assets/quality-map-slide.svg and docs/exploration-of-quality.md in that project.

Quality is how well the artifact satisfies its requirements. Evidence and provenance include sources, tests, calculations, reviews, and intermediate work. Assurance is the strength of the grounds for justified confidence. Assessability is how readily an expert can use that evidence to judge quality. Confidence is the evaluator’s belief about the work, influenced by expertise, prior beliefs, and biases. Calibration is how appropriately that confidence tracks actual quality.

Read the arrows as relationships in assessment, not a guarantee that good work automatically produces good evidence. Evidence can expose poor quality as well as support good quality. An extensive audit trail may provide assurance but remain difficult to assess. A polished summary may be easy to assess yet offer weak evidence.

The practical strategy is to preserve the evidence and make important claims easy for an expert to investigate. The aim is justified confidence in a particular artifact and its supporting evidence. This complements tests and evaluation sets for characterizing a system across many cases.

Operationalizing quality for AI output

Example: AI extracts structured metadata from research documentation.

Requirement	Observable criterion	Evaluator	Metric
Valid structure	Output conforms to schema	Validator	Pass/fail
Correct extraction	Values match the source	Reference labels / expert	Precision, recall, F1
Grounded	Each value has source support	Source review / calibrated judge	% supported
Complete	Required information is present	Reference labels / expert	Recall
Semantically valid	Relationships obey ontology	SHACL / rules	Violations
Useful	Researcher can use it without correction	Researcher review	Acceptance / error rate

Define the criterion first. Use the least subjective evaluator that can measure it.

Adapted from our prior discussion of AI output quality and its research-metadata example. This operationalizes the preceding diagram: evaluations produce evidence that can support assurance and calibrated confidence. Keep the source passages, reference labels, validation results, and review decisions accessible alongside the output to improve assessability.

Start with the task and its requirements, then choose observable criteria, evaluators, and metrics. Define acceptance thresholds before evaluation, according to the intended use and consequences of errors. The table is an evaluation design, not measured results.

For extraction, precision is the proportion of extracted items that are correct, and recall is the proportion of required reference items recovered correctly. F1 combines precision and recall. Specify the unit of evaluation, such as fields or entities, and the matching rule. For groundedness, report the fraction of extracted values with supporting source evidence. Source support does not by itself establish that a source is true. For usefulness, define acceptance and what counts as an error before researcher review.

Different dimensions require different evidence, and several evaluators may apply to the same output. A passing schema check establishes valid structure, not factual correctness. SHACL can validate graph constraints that encode selected ontology requirements. It does not establish every aspect of semantic correctness.

For open-ended work, use an explicit rubric covering observable properties such as major findings, attribution, preserved uncertainty, and conclusions that stay within the evidence. For subjective comparison, hide model identity and ask which output better satisfies a specific criterion. An LLM judge needs the task, source evidence, output, and rubric. Calibrate it against expert labels and check position, verbosity, and self-preference biases.

For a repeated system, examine pass rates, severe errors, variation across runs, task subtypes, and difficult edge cases. An average score can hide important failures. For an agent, also assess tool selection, evidence retrieval, unnecessary actions, failure detection, and recovery. These methods also support assessment of an individual artifact wherever its properties are testable.

Two productive patterns

The knowledge worker: use AI to produce a particular artifact.
The systems engineer: build AI into a capability that performs a class of tasks repeatedly.

Types of knowledge workers

Making — builders: How can this be built?
Directing — coordinators: What should be done, and in what order?
Deciding — leaders: Which outcomes matter, and which risks are acceptable?
Connecting — communicators: What will people understand, trust, value, or use?

Types of software engineers

Orientation	What they value	How they may need to grow
Craft	Mastery, correctness, elegance, authorship	Extend craft into architecture, evaluation, and stewardship
Systems	Interfaces, dependencies, constraints, failure modes	Develop domain and product judgment
Product	User needs and useful outcomes	Deepen evaluation and reliability skills
Operational	Observability, resilience, recovery	Help define autonomy, control, and acceptable failure earlier

Future jobs for software engineers

Directing AI-assisted production: specify work, guide agents, review results, and integrate changes.
Broader knowledge work: use AI across domains to produce analyses, designs, and products.
Engineering AI systems: build bounded, repeatable capabilities with evaluation and monitoring.
Conventional software engineering: pursue technical depth where security and direct control matter.

How does one choose a role?

Enjoyment: daily activities, immersion, creativity, satisfaction.
Mastery: strengths, technical depth, learning, direct authorship.
Impact: useful outcomes, purpose, values, people affected.
Compensation: income, benefits, financial needs, stability.
Lifestyle: hours, flexibility, location, workload, energy, family.
Community: colleagues, collaboration, belonging, mentorship.
Agency: autonomy, influence, ownership, responsibility.
Practical fit: available opportunities, experience, career stage, cost of retraining.

The role has changed… pick a new one

The old role can be dead even if the job title survives.
Change with AI: develop toward the work you want to do with it.
Or choose a different niche that better fits your interests and preferences.
Rebirth means choosing what comes next, rather than expecting the old job back.

Dealing with the change

Denial: This will not fundamentally change me.

Surrender: Nothing I choose will matter.

Agency:

I cannot know the outcome, but I can still participate deliberately in determining what happens next.

GitHub repository ↗ · Prep site & exploration notes ↗