Microsoft and OpenAI Extend to Custom AI Chips for Azure Maia and Cobalt

Microsoft’s partnership with OpenAI has moved decisively from software and cloud into silicon: Satya Nadella confirmed that Microsoft will be able to use OpenAI’s custom AI chip designs alongside its own in‑house efforts, giving Azure a legally backed pathway to incorporate OpenAI‑derived hardware ideas into Microsoft’s Maia and Cobalt initiatives.

Background​

For most of the last half‑decade Microsoft and OpenAI have operated an unusually close, commercially and technically integrated partnership that covered investment, cloud compute, and exclusive product tie‑ins. That relationship has been recast into a definitive agreement that extends Microsoft’s commercial and IP access windows for OpenAI models and — crucially — provides Microsoft rights to OpenAI’s hardware development work. The new arrangement also installs outside verification for any AGI declaration and reshapes exclusivity over compute provisioning.
This is not an abstract contractual tweak: the practical consequence is that Microsoft can legally inspect, adapt, and incorporate OpenAI hardware designs and system‑level networking ideas as inputs to its own Azure hardware roadmap — an accelerant to Microsoft’s stated objective of running “mainly Microsoft chips” in its AI data centers when and where it makes economic sense.

---

What Nadella Announced — The Practical Takeaway​

Satya Nadella’s remarks (given in a podcast released November 12) clarified how the revised IP deal will play out in operational terms: Microsoft has contractually backed access to OpenAI’s custom chip and networking designs, and will use those designs together with Microsoft’s internal hardware IP to accelerate its own silicon programs.
Key contractual windows also matter to the hardware story: Microsoft retains rights to OpenAI models through 2032 and extended research access through 2030 (or until an independent AGI verification panel deems AGI reached). Those legal timelines give Microsoft multi‑year runway to plan product integrations and to leverage OpenAI designs in negotiating foundry and components deals.
Put simply: Microsoft’s access is a strategic lever. It does not instantly replace external suppliers or mean immediate hyperscale deployment of OpenAI chips across Azure. But it reduces duplication of design effort, improves Microsoft’s negotiating posture with foundries and vendors, and gives the company an optional path to field differentiated inference hardware at scale.

---

Technical Snapshot: What We Know About OpenAI’s Chip Program​

Architecture and manufacturing node​

Reporting and internal briefings indicate OpenAI’s first custom accelerator uses a systolic array microarchitecture — a tiled, grid‑like arrangement of simple processing elements optimized for repeated matrix multiply‑accumulate operations, which are central to neural network inference. The chip is reportedly being developed for TSMC’s 3‑nanometer (N3) process.
Systolic arrays are a well‑bounded architectural choice for inference‑oriented ASICs: they yield high energy efficiency and predictable throughput for low‑precision tensor math, which makes them attractive for latency‑sensitive, high‑volume inference workloads. That said, they are not a drop‑in replacement for GPUs in large‑scale training, which still demands very high memory bandwidth, flexible precision modes, and mature software ecosystems.

Inference‑first focus, not training​

Crucially, OpenAI’s early part is aimed at inference — running trained models to serve outputs — not at the heavy, distributed training workloads that require massive FP16/FP8 throughput and huge memory capacity. Inference dominates operational costs for widely deployed services, so inference‑optimized silicon can deliver meaningful $/token savings once integrated at scale. But inference‑specialized chips typically can’t supplant GPUs used for frontier model training.

Timeline signals and production cadence​

Public reporting places initial mass production for OpenAI’s first custom part no earlier than 2026, with the chip being manufactured on TSMC’s 3‑nanometer node. That timeline is plausible but contingent: custom ASIC projects require multi‑stage validation (tape‑out, test silicon, yield ramp, packaging, system‑level integration), and foundry capacity for bleeding‑edge nodes remains constrained for hyperscale customers. Treat 2026 as a realistic but conditional milestone rather than a firm production date.

Systems work: Broadcom and networking​

OpenAI’s chip program reportedly includes systems and networking work with Broadcom, indicating the project is being conceived at rack and cluster scale rather than as a single die. That expands the value proposition for Microsoft: network topologies, switch fabrics, and packaging strategies are critical levers when deploying accelerators at hyperscale. Designs that pair compute tiles with optimized interconnect and aggregation can unlock latency, throughput, and utilization advantages at large scale.

---

How Microsoft Will Likely Use OpenAI Designs​

Microsoft’s stated approach is pragmatic and modular: it won’t necessarily manufacture an identical OpenAI part. Instead, Microsoft will evaluate microarchitectural blocks, power/clocking techniques, packaging and networking primitives from OpenAI’s designs and selectively adopt elements that complement Maia and Cobalt architectures and Azure’s datacenter systems.
Anticipated uses include:

• Adopting specific microarchitectural blocks (e.g., systolic array tiles) into Maia‑class accelerators for inference hotspots.
• Leveraging networking/topology designs for rack‑level orchestration and for optimized placement of inference clusters.
• Using OpenAI designs to improve negotiating leverage with foundries and external vendors when buying N3 wafers and packaging services.

This is a co‑design and selective assimilation strategy: the goal is to accelerate Microsoft’s internal roadmap and reduce duplicated NRE (non‑recurring engineering), not to create an immediate, wholesale substitution for Nvidia GPUs.

---

Why This Won’t Replace Nvidia Overnight​

Several structural realities make a near‑term, full replacement of Nvidia (or other GPU vendors) unlikely.

• Time to volume: Custom ASICs need multi‑stage validation and yield maturation. Initial production runs are often limited in volume and costly — Microsoft will not suddenly gain millions of datacenter units overnight.
• Software ecosystem inertia: Training toolchains (CUDA, cuDNN, mixed‑precision optimizers) and model libraries are heavily optimized for GPUs. Rewriting, validating, and optimizing these toolchains for new ASICs takes months or years and substantial engineering investment.
• Economics of ramp: Each new accelerator generation commonly involves hundreds of millions of dollars — chip NRE, packaging, board and rack redesign, and software stacks. Microsoft’s rights to OpenAI designs reduce duplication but do not eliminate the capital intensity of a hyperscale rollout.
• Different workload targets: The early OpenAI part is inference‑oriented. High‑end training for next‑generation foundation models will likely continue to depend on GPU ecosystems for the near term.

As a result, Azure’s likely near‑term posture is heterogeneous: continue to deploy Nvidia and AMD where they’re best, while gradually routing inference workloads to Maia and, when available and cost‑effective, OpenAI‑derived components.

---

Software and Portability: Where the Ecosystem Fits​

A mixed‑accelerator Azure will increase demand for portability layers and vendor‑neutral tooling. Microsoft already backs ONNX Runtime, which enables cross‑platform inference across multiple frameworks and hardware backends. Extending ONNX (and similar runtimes) to support new custom silicon will be a practical avenue for enabling model portability across Maia, OpenAI‑derived silicon, and existing GPU fleets.
Independent vendors and systems integrators are likely to seize the opportunity to build:

• Hardware‑agnostic compilers and runtime extensions that translate model graphs to the most efficient backend.
• Vendor‑neutral debugging, profiling, and observability layers that make cross‑hardware comparisons reproducible.
• ISV SDKs and tooling that reduce migration friction for enterprise customers moving from GPU‑only stacks to heterogeneous fabrics.

This software layer will determine whether mixed hardware becomes an operational advantage or a management headache for enterprise customers. Well‑engineered portability and orchestration tools are the linchpin for delivering predictable SLAs across mixed accelerators.

---

Economic and Strategic Implications​

Microsoft’s access to OpenAI designs delivers three immediate strategic advantages:

• Design leverage: Microsoft can learn from OpenAI’s experiments and re‑use proven microarchitectural blocks rather than reinventing them.
• Negotiation power: Having optionality across designs strengthens Microsoft’s hand with foundries and third‑party vendors during procurement negotiations.
• Product differentiation: Azure can offer specialized inference tiers optimized for latency, throughput, or privacy guarantees using a mix of Maia, OpenAI‑derived components, and commodity GPUs.

However, there are costs and trade‑offs. Rolling a new accelerator fleet at hyperscale entails:

• Large upfront capital and NRE expenses.
• Integration complexity at the system, rack, and network levels.
• Risk that expected $/inference savings are eroded by yield issues, slower than expected software optimization, or unanticipated integration bottlenecks.

In short: access to designs is strategic optionality, not a risk‑free shortcut to lower costs.

---

Risks, Fragilities, and What to Watch For​

Microsoft’s move unlocks optionality, but several risks deserve attention:

• Execution risk: Custom silicon rollouts at hyperscale are historically fraught — yield problems, packaging challenges, and late design bugs can delay rollouts and increase costs.
• Foundry constraints: TSMC N3 capacity is contested among hyperscalers. Any bottleneck or yield issue at N3 could push production timelines beyond 2026.
• Software ecosystem lag: If compilers, runtimes, and model optimizers lag hardware availability, real‑world performance will fall short of theoretical gains.
• Short‑term reliance on third parties: Even with design access, Microsoft will likely continue to lean on Nvidia and AMD for the highest‑end training and for immediate capacity needs. Expect a multi‑year hybrid posture.
• Unverifiable spec claims: Vendor TFLOPS and other marketing metrics should be treated cautiously until independent benchmarks are available; treat raw vendor numbers as indicative rather than definitive.

Any public statements about rapid substitution of Nvidia hardware or dramatic near‑term cost reductions should be viewed skeptically until reproducible, workload‑realistic benchmarks appear.

---

Practical Guidance for IT Leaders and Azure Customers​

Microsoft’s hardware strategy creates both opportunities and uncertainty for enterprise IT. Practical steps to prepare:

• Inventory and classify workloads by compute profile. Determine which workloads are latency‑sensitive inference vs. large‑scale training. Prioritize pilots for latency‑critical inference.
• Design for heterogeneity. Build abstraction layers today (ONNX, Triton) so workloads can be routed to the most cost‑effective backend without major code rewrites.
• Insist on reproducible benchmarks. For any cloud offering that advertises Maia or OpenAI‑derived silicon, require workload‑realistic A/B comparisons on latency, throughput, and $/inference.
• Negotiate clarity in procurement. If committing to large Azure contracts, ask for guarantees about hardware mix, placement policies, and SLAs for inference and training workloads. The mixed‑hardware future implies variability in the hardware serving a workload; contractual clarity matters.
• Partner with system integrators early. Enterprises that need turnkey migration from GPU stacks to mixed accelerators should engage integrators to handle benchmarking, sharding, and tuning. This becomes a technical and services opportunity.

---

Longer‑Term Outlook: What Success Looks Like​

If Microsoft successfully integrates OpenAI‑derived designs into Azure’s systems and ecosystems, several outcomes are plausible over the medium term:

• Meaningful inference cost reductions for high‑volume services, improving margins for Copilot, Bing, and large enterprise deployments.
• A stronger negotiating position with foundries and vendors, potentially easing supply constraints and lowering component costs over successive generations.
• A broad set of hardware‑agnostic tools and runtimes that allow model owners to treat the underlying accelerator as a managed commodity rather than a permanent lock‑in.

But realize that the transition is likely to be iterative. The near term will remain hybrid, and the pace of software and systems optimization will determine whether OpenAI‑derived silicon becomes a marginal complement or a central part of Azure’s cost and performance story.

---

Conclusion​

Microsoft’s right to use OpenAI’s custom chip designs is a significant strategic development that materially changes Azure’s optionality in the race for cost‑efficient, low‑latency inference at hyperscale. The revised agreement, with extended IP windows and access to hardware designs, gives Microsoft a legally backed path to accelerate its Maia and Cobalt programs and to orchestrate more sophisticated rack‑level topologies informed by OpenAI’s work.
That said, the pragmatic reality is a multi‑year, mixed‑hardware transition rather than an overnight revolution. Foundry constraints, software ecosystem inertia, capital intensity, and production risks mean Nvidia and other GPU suppliers will remain central to Azure’s high‑end training and many inference workloads for the foreseeable future. Microsoft’s access to OpenAI designs buys optionality, negotiation leverage, and a faster path to iterate on systems‑level ideas — but it does not eliminate the hard, expensive work of shipping a new accelerator fleet at hyperscale.
For IT leaders, the practical play is clear: prepare for heterogeneity, demand reproducible benchmarks, and design software portability now. The winners in this next phase will be the companies that treat hardware as one layer of a managed stack and prioritize cross‑platform runtimes, tooling, and observability that make heterogeneous accelerators an operational advantage rather than a vendor management headache.

---

Source: Tech in Asia https://www.techinasia.com/news/microsoft-to-use-openai-chip-designs-in-ai-hardware-push/

 

[ChatGPT]

Online betting is everywhere now — delivered through slick smartphone apps, integrated into live TV coverage, promoted by influencers, and built into the social fabric of major sporting events — and that reach comes with real, measurable benefits to operators and real, measurable harms to people who wager beyond their means. As the Associated Press recently noted, roughly 14% of U.S. adults report betting on professional or college sports online either frequently or occasionally, and expert voices warn that convenience and velocity make online wagering a qualitatively different risk than the old trip to a casino or a corner bookie.

Background​

Online gambling and sports betting have transformed from niche activities into mainstream consumer services in a matter of years. That transition has several visible consequences: record industry revenues, expanding product sets (from micro‑bets and in‑play props to prediction markets), and a sharp rise in public scrutiny because of player harm and a string of high‑profile integrity scandals. In the U.S., commercial casinos posted a record year in 2023, winning an estimated $66.5 billion — with internet gambling and sports betting representing fast‑growing slices of that total. At the same time the business is booming, the lines between video games, social apps, and gambling — whether via loot‑box mechanics or tokenized rewards — have blurred considerably. Community conversations and tech‑industry thread archives show how monetization techniques borrowed from casinos are increasingly normalized in entertainment software, heightening the urgency of consumer protection and regulation.

---

The landscape today: growth, harm, and headline risk​

Rapid commercial expansion​

• Online sports betting and casino sites have delivered explosive topline growth for major operators. FanDuel and DraftKings, among others, have scaled massively in the United States after legal expansion following the 2018 Supreme Court decision that opened the door to state‑level legalization. This expansion is a principal driver behind the sector’s revenue records.
• Analysts and investigative reporting show operators exporting products and marketing strategies into newer markets with fewer safeguards, raising questions about whether growth is proceeding at the expense of consumer protection.

Rising public‑health and social costs​

• Large scale surveys and academic studies, particularly in the UK, have documented a notable uptick in gambling harm over the past few years. One respected survey suggests that harm from problem gambling could be many times larger than earlier estimates — with young men, disadvantaged communities, and in‑play sports bettors disproportionately affected. These findings have shifted public debate and intensified calls for regulation.
• Those harms show up in many ways: financial distress, relationship breakdown, mental‑health problems, and, tragically, suicide in the most acute cases. Public‑health advocates stress that the speed and anonymity of online platforms can accelerate harm compared with in‑person experiences.

Integrity and scandal​

• Betting integrity groups and industry monitors continue to detect suspicious activity linked to online markets. Sportradar’s monitoring concluded that the overwhelming majority of sport events remain untainted, but the system flagged hundreds to a few thousand suspicious matches globally each year — and those anomalies often cluster around lower‑tier competitions and micro‑bet lines that are highly vulnerable to manipulation.
• The International Betting Integrity Association (IBIA) and other watchdogs have reported increases in suspicious betting alerts at times, highlighting the fragility of specific bet types (for example, pitch‑level or single‑play props) and the ease with which small amounts of inside information can sway micro‑markets. Recent high‑profile cases — involving players, coaches, officials, and support staff — have made the headlines and prompted calls for reform.

---

Why online gambling is different — and why that matters​

Velocity, accessibility, and frictionless loss​

Digital wagering replaces friction with immediacy. A few taps can turn curiosity into a bet, and rapid, in‑play markets allow dozens or hundreds of wagers within a single match. That speed works in favor of platforms (increased activity and margin) and against bettors (higher potential losses and impaired reflection). Public‑health specialists emphasize that ease of access increases the risk of problem behaviour because time to reflect or stop is compressed.

Product design and behavioural cues​

Modern platforms employ features that encourage repeat play and higher stakes: push notifications timed to events, loyalty or VIP programs, personalized offers based on account activity, and “boosts” that make bets feel more valuable. Similar techniques are visible in mobile games and other apps, blurring the user’s sense of risk and reward. Several community analyses of game monetization confirm the ethical risk when casino‑style mechanics are embedded in entertainment products.

Narrow markets, wide exposure​

Micro‑bets and player props drastically reduce the number of actors needed to influence an outcome. Where season‑long outcomes require broad collusion, a pitch‑level prop or a specific statistic can be materially affected by a single individual’s action or non‑action. Betting integrity firms have warned that these thin markets are disproportionately targeted in manipulation schemes.

---

Practical safeguards operators offer — and where they fall short​

Standard responsible‑gambling tools​

Most licensed operators now provide a toolkit of user controls:

• Deposit limits (daily/weekly/monthly caps)
• Loss limits and wager limits
• Session time limits and reality checks
• Time‑out (short break) and self‑exclusion features
• Option to restrict deposit methods or remove payment rails temporarily

These tools are widely available at major operators and are central elements of corporate compliance programs. DraftKings, FanDuel and others publish responsible‑gaming pages describing these exact options.

Third‑party blocking and national self‑exclusion​

• In the UK, GAMSTOP, the national online self‑exclusion service, has seen record registration growth — hundreds of thousands of users have signed up to ban themselves from licensed sites, and registrations among under‑25s have risen notably. This reflects both growing awareness and increasing incidence of harm in younger cohorts.
• Consumer‑facing blocking tools (like Gamban and BetBlocker) exist to help people who want to cut access across devices and operators; such products are a practical complement to operator tools but are not a cure and depend on user buy‑in and technical coverage.

Limits of company controls​

Operators’ safeguards are sometimes optional, opt‑out, or implemented with default settings that favour engagement. Investigations have shown that companies operating in multiple jurisdictions will sometimes apply more conservative protections where regulators demand them (for example, in the UK) and looser settings where regulation is lighter, raising questions about whether consumer protection is being driven by compliance or by commercial calculation.

---

How individuals can reduce risk: concrete steps that work​

The following practical measures are framed for people who are considering online gambling or already participate occasionally. They are not medical advice but represent widely recommended harm‑reduction practices.

1. Treat gambling as entertainment — budget it like entertainment​

• Set a strict monthly entertainment budget and never use money needed for rent, bills, loan repayments or groceries.
• Decide in advance the small portion of that entertainment budget you might spend on gambling and make that a non‑negotiable cap.

2. Use platform controls immediately — and make them conservative​

• Activate deposit limits, loss limits, and session time limits as soon as you open an account. Lower limits take effect immediately in most systems; increases often require a cooling‑off period.
• Use time‑out or self‑exclusion options if you feel tempted to increase stakes or chase losses. Where available, national self‑exclusion schemes like GAMSTOP add an extra layer of protection across licensed platforms.

3. Remove the frictionless payment methods​

• Avoid storing payment cards with betting apps, and limit payment options where the platform allows it. Some services let you block specific deposit methods to reduce impulsive funding.
• Consider using pre‑ funded accounts or separate debit cards that are topped up from your entertainment budget only.

4. Use blocking software and device controls​

• Install device‑level blocking tools such as Gamban, BetBlocker or browser extensions that limit or block access to gambling sites. These tools are not perfect but are effective when combined with behavioural commitments.
• On phones, use app‑time controls or parental‑control suites to enforce session limits and lockouts.

5. Monitor behaviour and set external alarms​

• Use reality checks and periodic financial reviews. If you notice time spent or money lost increasing, escalate immediately: tighten limits, take a break, and consider self‑exclusion.
• Ask a trusted friend or family member to periodically review account statements with you or hold the passwords in escrow (mutual accountability can be a powerful control).

6. Don’t rely on “systems” or tips; accept the house edge​

• Betting systems that promise “sure” returns are mathematically unsound for the market as a whole. Treat any betting system as entertainment — not an investment strategy.

7. Seek help early​

• If gambling is causing stress, debt, relationship problems, or anxiety, reach out to professional support early. National helplines, gambling‑treatment charities and clinical services exist in many jurisdictions and are often free or low‑cost.

---

For policymakers and platform designers: what needs to change​

The problem is not just individual choices; product design and regulatory frameworks drive both risk and reward. The following are policy and product interventions that would materially reduce harm if implemented well.

• Default‑on protections. Make harm‑reduction tools default settings rather than optional opt‑ins. Default‑on deposit caps, reality checks, and lower default bet sizes would reduce impulsive loss.
• Tighter rules for micro‑bets and props. Restrict or require higher scrutiny on sub‑event markets where a single actor can meaningfully affect outcomes.
• Transparency and auditing. Require companies to publish anonymized account‑level data to regulators for integrity monitoring and independent audits of algorithms that target consumers.
• Cooling‑off mandates for increases. Require minimum cooling‑off periods for any increase in deposit or loss limits to prevent impulsive escalations.
• Universal self‑exclusion registries. Encourage interoperable national registries where feasible, paired with robust identity‑protection protocol design.

These ideas are not new, but the pace of product innovation has outstripped the regulatory response in many jurisdictions. Investigations show operators sometimes vary their degree of consumer protection depending on local enforcement, a practice that undercuts claims of uniform responsible gambling commitment.

---

Integrity risks — why regulation matters for sports and consumers​

Recent investigations and industry reports show how integrity risks concentrate around thin markets and non‑standard bet types:

• Monitoring groups find most sporting events remain untainted, but even a small absolute number of manipulated matches can produce outsized harm given the volume of money moving through modern betting markets. Sportradar’s analysis and industry alerts emphasize the continued need for robust, account‑level data sharing between operators and sporting bodies to detect manipulation early.
• Suspicious betting alerts, flagged by entities like IBIA, have shown spikes in certain periods and markets. Those alerts are a leading indicator: they may presage criminal investigations, sanctions, and reputational damage for leagues and operators if not acted upon swiftly. Regulators and operators need cooperative, cross‑border surveillance to keep pace.

---

Strengths in the current system — why some safeguards work​

• Operator toolkits are real and improving. Deposit limits, time‑outs, and self‑exclusion features are practical and demonstrably useful when activated. Many operators now build these options into account settings and make them easy to find.
• National registries and blocking tools are scaling. Services like GAMSTOP have grown to half‑a‑million registrants and show that people will use centralized self‑exclusion when it is straightforward and well‑promoted.
• Integrity monitoring works when data is shared. Sportradar and other firms show that account‑level market data combined with AI detection produces actionable alerts that have stopped manipulation in many cases. This demonstrates that technical solutions scale when the industry commits to transparency and collaboration.

---

Risks, weaknesses, and unanswered questions​

• Optionality and defaults. Too many protections remain optional; making them default would reduce harm substantially but invites industry resistance and legal questions about paternalism.
• Global inconsistencies. Operators often deploy stronger safeguards in jurisdictions with stricter enforcement and less in newer or less regulated markets. This uneven approach undermines claims of corporate social responsibility.
• Data privacy vs. detection. Effective integrity monitoring and harm detection require granular data. Balancing privacy and utility remains a thorny technical and ethical issue.
• Unverifiable claims. Industry assurances about the effectiveness of VIP monitoring, targeted limits, and remediation are sometimes opaque; independent verification and cross‑jurisdiction audits are still a work in progress. Where companies make internal claims about harm reduction, those assertions must be scrutinized by regulators and researchers.

When public reporting or company claims cannot be independently verified, those claims should be flagged as provisional. Investigative reporting has already shown cases where declared protections were not implemented or were changed quietly as operators expanded into new markets. Readers should treat unilateral corporate statements about “best practices” with cautious scrutiny unless backed by verifiable, third‑party audit evidence.

---

A practical checklist to reduce personal risk (quick reference)​

• Set deposit limits (daily/weekly/monthly) immediately.
• Turn on session timers and reality checks.
• Avoid storing payment cards with betting apps; use a separate account/card for entertainment spending.
• Install third‑party blocking software and enable device‑level restrictions.
• If gambling becomes a source of stress or debt, use time‑out or self‑exclusion and seek professional help early.

---

Conclusion​

Online gambling today sits at an uneasy crossroads: it is profitable, technologically innovative, and widely available, yet it amplifies familiar human vulnerabilities through design choices that privilege speed, engagement, and monetization. The good news is that tools exist — deposit limits, time‑outs, self‑exclusion registries, and blocking apps — and independent monitoring systems can and do detect manipulation and suspicious activity. The bad news is that these protections remain inconsistently applied, sometimes optional by design, and are outpaced by product innovation.
A safer online gambling environment is possible, but it will require three things to align: sensible regulation with enforceable defaults, operators committing to transparent, verifiable protections across all markets, and consumers taking concrete steps — including conservative use of platform controls and blocking tools — to protect their finances and well‑being. Until those pieces fit together more reliably, the simplest and most effective strategy for an individual remains the same: treat gambling as entertainment, set strict financial and time boundaries, and use the technical and support tools that are available to enforce them.

---

Source: Newswav Online gambling is growing in popularity. Here's how to avoid its biggest pitfalls