Markets Aren't Buying the "Blame AI" Layoffs

---

TL;DR

Monday.com cut 20% of its workforce (over 600 people), framing it as a pivot to an “AI-driven growth strategy” — but an FT analysis found that companies citing AI as a reason for layoffs underperformed the Nasdaq by nearly 10% on average in the 30 trading days following the announcement, suggesting the market isn’t taking these explanations at face value. US tech companies have cut roughly 140,000 jobs so far this year, even as Anthropic and OpenAI rapidly expand hiring and Meta reassigns 7,000 employees to AI-related roles — a mix of net reductions and reshuffling rather than a single clean story. On arXiv, new research spans agent training infrastructure (OpenForgeRL) and VLM reliability issues — self-distillation, cross-modality inconsistency, and the structure of sycophancy — all with direct implications for operating cost and trustworthiness.

Main story: the market itself is skeptical of “AI-driven layoffs” as an explanation

Tel Aviv-based Monday.com disclosed in an SEC filing that it is cutting about 20% of its workforce (over 600 people). Co-founder Eran Zinman said in an internal LinkedIn memo that the move is “not about cost cutting or replacing people with AI,” positioning it instead as part of the “AI-first” pivot the company announced roughly a year ago. Monday.com expects $45–55 million in net restructuring charges while projecting revenue growth of up to 20% year-over-year for full-year 2026. (TechCrunch)

According to a Financial Times analysis, US tech companies have already cut about 140,000 jobs this year, with Amazon, Oracle, Meta, and Microsoft alone accounting for roughly 50,000 of those. The same companies are simultaneously pouring hundreds of billions of dollars into data center expansion. The FT found that companies citing AI as a reason for layoffs underperformed the Nasdaq by nearly 10% on average in the 30 trading days following the announcement — a finding it says suggests the market isn’t taking companies’ explanations at face value.

Technical read

What these stories share is a near-total absence of technical evidence for what AI actually replaced. Monday.com’s Zinman explicitly said the goal wasn’t “replacing people with AI,” and Microsoft similarly framed its role cuts as not being AI replacement, while offering only the vague claim that “AI is changing how work gets done.” In other words, what this reporting establishes is that the word “AI” is being invoked in layoff explanations — not that any specific job function was actually replaced by AI. This isn’t a question of demonstrated capability but of corporate messaging, and it’s a case where the GIGO principle (questioning input quality) should be applied to companies’ AI claims themselves.

Business read

• Unit economics/cost structure: Monday.com is pairing a 20% headcount cut with a projected revenue increase of up to 20% year-over-year — a bullish scenario combining cost restructuring with continued growth. But the FT’s stock price data suggests investors aren’t buying this narrative (cost cuts = AI-driven efficiency).
• Net reduction vs. reshuffling: Not all “AI layoffs” are the same. Meta is placing 7,000 people into AI-related roles while cutting a separate 8,000, and IBM tripled new-grad hiring for AI and hybrid-cloud roles alongside its own recent cuts. Microsoft’s 4,800-person cut in its gaming division (Xbox), by contrast, appears more tied to resetting the integration of its $75 billion Activision Blizzard acquisition — the sourcing doesn’t establish a direct AI causal link there.
• Talent flows: Anthropic and OpenAI are hiring rapidly, absorbing at least some of the talent shed elsewhere. This offers circumstantial evidence that the center of gravity in the labor market is shifting from companies applying AI to the AI foundation-model companies themselves.
• The market as an evaluation source: The FT’s stock-reaction data models a useful methodology in itself — rather than taking a company’s self-reported narrative (“it’s AI-driven, therefore legitimate”) at face value, cross-check it against an independent external signal, in this case stock price.

Contrarian take / what might be missed

The “AI layoffs” framing itself may be a convenient single causal story for both companies and the press. In reality, multiple motives — restructuring, market conditions, capital reallocation toward AI infrastructure spending — likely coexist. Even the source (the FT analysis) doesn’t assert that stock underperformance is definitively caused by skepticism toward AI explanations; it uses language like “suggests.” So this stock reaction should be read as correlation, not proof of “market judgment on AI adoption.”

Implications and positioning

Don’t use big companies’ “AI-driven” framing as evidence to inform your own AI investment decisions. What’s actually disclosed is only that the word “AI” is being used — not the specifics or the effectiveness of any automation. The methodological lesson is more useful: when evaluating another company’s AI results, prioritize independent external signals (stock reaction, customer churn rate, etc.) over self-reported claims (press releases, IR materials). Signal/noise: the fact of the layoffs themselves and the FT’s aggregate data are high-signal. But the causal claim that “AI is the technical cause of these job cuts” remains a largely unsubstantiated narrative at this point — low confidence.

Other notable items

OpenForgeRL: how much agent training infrastructure should you build in-house?

A framework for end-to-end reinforcement learning training of agents that use complex reasoning harnesses like Claude Code, Codex, or OpenClaw — using the actual harness and actual environment, unmodified. A lightweight proxy records the harness’s model calls and converts them into training data for existing RL infrastructure (e.g., veRL), while a Kubernetes orchestrator runs each rollout in an isolated remote container. Using only a few hundred to a few thousand tasks, the tool/claw-based agent OpenForgeClaw reportedly outperforms similarly-sized existing open models on nearly all benchmarks in ClawEval (31.7 pass^3 / 55.9 pass@3) and QwenClawBench (33.7); the GUI-operation agent OpenForgeGUI similarly leads on OSWorld-Verified (37.7), Online-Mind2Web (63.0), and WebVoyager (72.3), even matching or exceeding models several times its size on GUI-operation tasks. (arXiv)

Technical read: Being able to run RL directly on the actual production harness and environment — rather than a simplified training environment — closes the gap between agent training and real-world deployment. Decoupling training from inference means RL can be applied without rewriting existing reasoning harnesses. Business read: This is exactly the kind of area where “investing in proprietary infrastructure risks being absorbed by an official stack later.” If Anthropic or OpenAI start offering native fine-tuning on top of their own harnesses, the value of third-party training infrastructure like this erodes.

Implications and positioning: If you’re building agent products, avoid investing in rebuilding this kind of training infrastructure yourself, and instead watch whether official providers start offering native fine-tuning on their harnesses. What remains as differentiation isn’t the training infrastructure itself, but your own task- and environment-specific data. Signal is medium-to-high (addresses a real infrastructure problem, backed by benchmarks), but platform-absorption risk is high.

Visual Contrastive Self-Distillation: training VLMs without an external teacher model

A method that distills a student model using the token-level difference between two distributions generated by an EMA teacher model under the same prompt and prefix — one conditioned on the original image, one conditioned on a version with image content erased. It requires no external teacher model or privileged ground-truth data. On Qwen3-VL, the average score across 7 benchmarks improved from 62.27% to 67.04% for the 2B model, 71.30% to 73.16% for the 4B model, and 72.51% to 76.26% for the 8B model. (arXiv)

Technical read: Because distillation no longer requires separate API calls to a large external teacher model, the underlying cost structure of training changes. The improvement being consistent across model sizes is a positive sign for reproducibility. Business read: For teams looking to fine-tune small VLMs in-house for products, avoiding reliance on a large external teacher model could reduce API costs correspondingly.

Implications and positioning: If you want to cut training costs for small VLMs, it’s worth investigating and adapting this kind of self-distillation approach. That said, validation so far is limited to the ViRL39K dataset and the Qwen model family — reproduction on other model families is still pending. Signal is moderate.

MIRROR: the VLM flaw where the same content gets different answers depending on “diagram” vs. “text”

A paper demonstrating that VLMs give different correct/incorrect results depending on presentation format — for problems like geometry that have equivalent text-only, diagram-only, and text+diagram representations. The paper finds cases where a model can solve a problem in text but not in diagram form, and vice versa. It compares performance across presentation formats and proposes MIRROR, a reinforcement learning method that uses the best-performing format as a “teacher” to train the others via a reverse-KL objective, improving both consistency and accuracy on geometry benchmarks. (arXiv)

Technical read: This isn’t a ceiling-on-capability problem — it’s a reliability/GIGO-adjacent flaw where the model “knows the answer but can’t retrieve it when the presentation changes.” Business read: This points to a concrete risk for VLM products that handle both documents and diagrams (invoice/chart reading, QA over documents with figures, etc.) — namely, vulnerability when the same fact is presented in slightly different visual representations.

Implications and positioning: If you’re running a VLM product that handles both text and diagrams, don’t trust a single benchmark score — add a test that asks the same question in both diagram and text form. As a diagnostic, the signal is strong, but MIRROR as a solution has only been validated in the geometry domain; generalization to other domains is unconfirmed.

Beyond Sycophancy: model “sycophancy” isn’t a single-dimensional flaw

A study analyzing LLM sycophancy not as a simple flaw but as a structured judgment-revision process shaped by three axes: (1) the distance between the input opinion and the model’s own initial judgment, (2) the attribution of the opinion’s source, and (3) the structure of the group that holds that opinion. The results parallel phenomena known from human social psychology. (arXiv)

Technical read: This isn’t a new training method or benchmark result — it’s a diagnostic reframing of sycophancy as a phenomenon. Business read: For agent products that provide advice or decision support to teams, the finding that a model’s judgment revision can vary based on the opinion’s source and the structure of the group backing it points to concrete risks like consensus manipulation or amplified groupthink.

Implications and positioning: If you’re running a decision-support agent, it’s worth specifically testing how the model responds to pushback framed as collective (“the whole team thinks this”). The diagnosis itself is concrete and testable, but no training countermeasure has been proposed yet — confidence remains at the level of an academic suggestion.

GraphVid: shifting multi-object control in video generation from “trajectories” to “graphs”

When generating video involving multiple interacting objects, conventional trajectory-based control requires users to draw precise trajectories for each object, which breaks down as object counts increase or occlusion occurs. This paper proposes GraphVid, an image-to-video generation model that lets users specify relationships between objects via a structured interaction graph, along with GraphVid-Bench, a large-scale dataset focused on interactions. Using less training data and fewer parameters than prior motion-control methods, it improves FID by up to 39.9% and FVD by 37.6% compared with Motion-I2V, and also improves PSNR (9.87 → 15.98) and SSIM (0.38 → 0.61). (arXiv)

Technical read: Replacing the manual bottleneck of trajectory drawing with structured graph input is a concrete design change, and outperforming Motion-I2V on FID/FVD with less training data and fewer parameters than prior motion-control methods also demonstrates lower training cost. Business read: Graph-based interaction specification is a more API-friendly, template-friendly interface than trajectory drawing, well-suited to product designs that don’t require experts to manually draw trajectories. This could inform design decisions for video generation tools involving multiple interacting objects — advertising, product demos, game asset production, and similar use cases.

Implications and positioning: If you’re considering a video generation tool for scenes with multiple interacting objects, prioritize a structured-graph interface over a trajectory-drawing one. The design is more accessible to non-experts, but the benchmark advantages come from a single paper and await third-party reproduction.

Worth trying this week / hype to ignore

• Worth trying this week: When evaluating your own agent products, look at the pass@k-style evaluation approach OpenForgeRL uses via ClawEval as a template for reproducibility and cost-checking on your own tasks — no need to build the training infrastructure itself.
• Hype to ignore: Big companies’ “AI-driven layoffs” framing. Absent disclosed specifics about what was actually automated, this shouldn’t inform your own AI investment decisions — and the FT’s finding that the market itself isn’t taking this narrative at face value backs that up.